Make parsing even faster, and use less memory

[gpittarelli]

Further refinement to #107. cpu+mem profiler in emacs still showed a global scope completion with javascript-typescript-langserver chewing through 119MB and GC was taking up 28% of the runtime. Profiling still pointed towards lsp--parser-read. I suspected constantly concat'ing chunks was making a ton of extra allocs; eg we'd get a 4k chunk initially, then we'd alloc a new 8k chunk when (to hold leftover 4k + new chunk), then 12k, 16, ... compounding until we reached the full megabyte response.

This change preallocates a single big string of the total Content-Length size, then fills it as we go. This cuts memory usage to around 29M and GC cpu time to ~23%.

Note: trusting the server to not set a malicious Content-Length...

As an aside, json-read-from-string is now the the single biggest CPU bottleneck, followed by GC and then various internal emacs completion stuff. So I guess we're good for now?

Number are from running a test script with emacs -q --script. Profiles are in the zip (open with profiler-find-profile)
profiles.zip

[vibhavp]

Thanks!