Fix slow dot in numba

[ricardoV94]

Closes #1418

I confirmed with perf that numba is calling gemv, the problem was the castings astype to handle mixed or discrete dtypes always force a copy even when the cast dtype is the same as the input (discussed in numba/numba#10085)

We should check the other functions that make use of int_to_float_fn (mostly linalg stuff), and adapt accordingly. This PR is just to fix the dot case.

Benchmark results

Before
------------------------------------------------------------------------------------------------- benchmark: 3 tests -------------------------------------------------------------------------------------------------
Name (time in us)                              Min                   Max                Mean             StdDev              Median                IQR            Outliers  OPS (Kops/s)            Rounds  Iterations
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_mat_vec_dot_performance[float32]      48.5510 (1.0)      1,294.7070 (3.11)      61.3219 (1.0)      14.6329 (1.0)       60.6585 (1.0)       2.5140 (1.0)      231;2385       16.3074 (1.0)        9052           1
test_mat_vec_dot_performance[mixed]       150.2420 (3.09)       514.2840 (1.23)     172.9162 (2.82)     34.0547 (2.33)     159.1785 (2.62)     19.9170 (7.92)      298;300        5.7831 (0.35)       3162           1
test_mat_vec_dot_performance[float64]     158.5070 (3.26)       416.6410 (1.0)      181.0844 (2.95)     33.2207 (2.27)     165.9410 (2.74)     23.4185 (9.32)        73;59        5.5223 (0.34)        725           1
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

After
----------------------------------------------------------------------------------------------- benchmark: 3 tests ----------------------------------------------------------------------------------------------
Name (time in us)                             Min                 Max               Mean             StdDev             Median                IQR            Outliers  OPS (Kops/s)            Rounds  Iterations
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_mat_vec_dot_performance[float32]     23.5340 (1.0)       69.0690 (1.0)      30.1434 (1.0)       3.2675 (1.0)      29.1650 (1.0)       0.5820 (1.0)     1465;3237       33.1748 (1.0)       16409           1
test_mat_vec_dot_performance[float64]     40.3260 (1.71)     159.3380 (2.31)     51.0019 (1.69)      8.8649 (2.71)     51.2260 (1.76)      9.6455 (16.57)    1587;272       19.6071 (0.59)       6019           1
test_mat_vec_dot_performance[mixed]       40.5660 (1.72)     180.4480 (2.61)     53.0003 (1.76)     16.4452 (5.03)     50.8750 (1.74)     11.4110 (19.61)     923;845       18.8678 (0.57)       9525           1
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

It now matches the C-backend performance in my machine

---

📚 Documentation preview 📚: https://pytensor--1426.org.readthedocs.build/en/1426/

[ricardoV94]

The old check doesn't work or stopped working at some point:

assert (
    isinstance(np.dtype("float64"), np.floating)
    or isinstance(np.float64, np.floating) 
)  # fails

[ricardoV94]

We don't use this function anymore in dot, but it's used elsewhere

[ricardoV94]

I need these branches otherwise, I get some failed to unify during numba compilation.
It doesn't like the pattern:

if x.dtype != dot_dtype:
  x = x.astype(dot_dtype)

When it's actually needed. Can be simplified once astype(copy=False) is implemented in numba

[codecov]

Codecov Report

Attention: Patch coverage is 81.25000% with 6 lines in your changes missing coverage. Please review.

Project coverage is 82.11%. Comparing base (261aaf3) to head (5f7a740).
Report is 10 commits behind head on main.

Files with missing lines	Patch %	Lines
pytensor/link/numba/dispatch/basic.py	81.25%	4 Missing and 2 partials ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1426   +/-   ##
=======================================
  Coverage   82.11%   82.11%           
=======================================
  Files         211      211           
  Lines       49686    49743   +57     
  Branches     8813     8824   +11     
=======================================
+ Hits        40798    40847   +49     
- Misses       6710     6715    +5     
- Partials     2178     2181    +3     

Files with missing lines	Coverage Δ
pytensor/link/numba/dispatch/basic.py	79.08% <81.25%> (-0.46%)	⬇️

... and 6 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[aseyboldt]

Looks good :-)