evaluate using cattrs + orjson instead of custom serialization/deserialization

seems like as of cachew v0.13.0, on test_many, cattrs + orjson is

- 31% faster for writing
- 17% faster for reading

Note that this includes other overhead like writing to the DB, so actual serialization/deserialization is even faster than that

performance stats:

original version:

```
src/cachew/tests/test_cachew.py::test_many[1000000-False] [INFO    2023-09-10 23:22:55,188 cachew __init__.py:1141] cachew.tests.test_cachew:test_many.<locals>.iter_data: wrote   1000000 objects to   cachew (sqlite /tmp/pytest-of-adhoc/pytest-74/test_many_1000000_False_0/test_many)
test_many: initial write to cache took 7.2s
test_many: cache size is 23.06048Mb
[INFO    2023-09-10 23:22:55,306 cachew __init__.py:1005] cachew.tests.test_cachew:test_many.<locals>.iter_data: loading 1000000 objects from cachew (sqlite /tmp/pytest-of-adhoc/pytest-74/test_many_1000000_False_0/test_many)
test_many: reading from cache took 3.7s
[INFO    2023-09-10 23:22:58,995 cachew __init__.py:1005] cachew.tests.test_cachew:test_many.<locals>.iter_data: loading 1000000 objects from cachew (sqlite /tmp/pytest-of-adhoc/pytest-74/test_many_1000000_False_0/test_many)
PASSED
```

with cattrs + orjson

```
src/cachew/tests/test_cachew.py::test_many[1000000-False] [INFO    2023-09-10 23:23:35,704 cachew __init__.py:1055] cachew.tests.test_cachew:test_many.<locals>.iter_data: wrote   1000000 objects to   cachew (sqlite /tmp/pytest-of-adhoc/pytest-75/test_many_1000000_False_0/test_many)
test_many: initial write to cache took 4.9s
test_many: cache size is 72.904704Mb
[INFO    2023-09-10 23:23:36,048 cachew __init__.py:916 ] cachew.tests.test_cachew:test_many.<locals>.iter_data: loading 1000000 objects from cachew (sqlite /tmp/pytest-of-adhoc/pytest-75/test_many_1000000_False_0/test_many)
test_many: reading from cache took 3.1s
[INFO    2023-09-10 23:23:39,114 cachew __init__.py:916 ] cachew.tests.test_cachew:test_many.<locals>.iter_data: loading 1000000 objects from cachew (sqlite /tmp/pytest-of-adhoc/pytest-75/test_many_1000000_False_0/test_many)
PASSED
```

This will break many other things at the moment though, so can't use it straightaway

Author: karlicoss

src/cachew/__init__.py

@@ -17,6 +17,8 @@
 import dataclasses
 import warnings
 
+import cattrs
+import orjson
 
 import appdirs
 
@@ -562,9 +564,10 @@ def do_begin(conn):
 
         self.binder = NTBinder.make(tp=cls)
         # actual cache
-        self.table_cache     = Table('cache'    , self.meta, *self.binder.columns)
+        # FIXME change table definition
+        self.table_cache     = Table('cache'    , self.meta, Column('data', sqlalchemy.String))
         # temporary table, we use it to insert and then (atomically?) rename to the above table at the very end
-        self.table_cache_tmp = Table('cache_tmp', self.meta, *self.binder.columns)
+        self.table_cache_tmp = Table('cache_tmp', self.meta, Column('data', sqlalchemy.String))
 
     def __enter__(self) -> 'DbHelper':
         return self
@@ -882,7 +885,7 @@ def composite_hash(self, *args, **kwargs) -> Dict[str, Any]:
         }
         kwargs = {**defaults, **kwargs}
         binder = NTBinder.make(tp=self.cls_)
-        schema = str(binder.columns) # todo not super nice, but works fine for now
+        schema = str('FIXME') # todo not super nice, but works fine for now
         hash_parts = {
             'cachew'      : CACHEW_VERSION,
             'schema'      : schema,
@@ -993,11 +996,14 @@ def cachew_wrapper(
 
             logger.debug('old hash: %s', old_hash)
 
+            converter = cattrs.Converter()
 
             def cached_items():
                 rows = conn.execute(table_cache.select())
-                for row in rows:
-                    yield binder.from_row(row)
+                for (js,) in rows:
+                    xx = orjson.loads(js)
+                    rr = converter.structure(xx, binder.type_)
+                    yield rr
 
             if new_hash == old_hash:
                 logger.debug('hash matched: loading from cache')
@@ -1107,9 +1113,11 @@ def flush() -> None:
                         dict(zip(column_names, row))
                         for row in chunk
                     ]
-                    conn.execute(insert_into_table_cache_tmp, chunk_dict)
+                    conn.execute(insert_into_table_cache_tmp, [{'data': c} for c in chunk])
                     chunk = []
 
+            converter = cattrs.Converter()
+
             total_objects = 0
             for d in datas:
                 try:
@@ -1118,8 +1126,11 @@ def flush() -> None:
                 except GeneratorExit:
                     early_exit = True
                     return
-                  
-                chunk.append(binder.to_row(d))
+
+
+                js = converter.unstructure(d, binder.type_)
+                js = orjson.dumps(js) # .decode('utf8') # TODO just dump as bytes?? decoding takes time
+                chunk.append(js)
                 if len(chunk) >= chunk_by:
                     flush()
             flush()

src/cachew/tests/test_cachew.py

@@ -78,7 +78,8 @@ def test_ntbinder_primitive(tp, val) -> None:
     assert vv == val
 
 
-class UUU(NamedTuple):
+@dataclass
+class UUU:
     xx: int
     yy: int
 
@@ -285,7 +286,10 @@ def fun2() -> Iterable[Union[UGood, UBad]]:
         list(fun2())
 
 
-class TE2(NamedTuple):
+# NOTE: had to change this from NamedTuple, since cattrs doesn't seem to handle NamedTuple correctly atm
+# performance for current cachew is same for dataclass/NamedTuple, so doesn't impact benchmarks
+@dataclass
+class TE2:
     value: int
     uuu: UUU
     value2: int