perf(python): optimize pystr deserialize perf

[chaokunyang]

What does this PR do?

This PR implemented an optimized version of PyUnicode_FromUCS1/Fury_PyUnicode_FromUCS2 for faster performance by :

• replace max char check using SIMD
• Cast ucs2 array to ucs1 array by SIMD

Related issues

Does this PR introduce any user-facing change?

• Does this PR introduce any public API change?
• Does this PR introduce any binary protocol compatibility change?

Benchmark

[chaokunyang]

cc @penguin-wwy @pandalee99 @theweipeng

[pandalee99]

This code is very efficient,very nice!

maybe we can optimize the repetitive code.

  // Handle remaining elements
  for (; i < length; i++) {
    if (arr[i] > max_sse) {
      max_sse = arr[i];
    }

It's just the way it's written. It's nothing serious.

[chaokunyang]

If i use PyUnicode_DecodeLatin1 directly here, It's faster in macos, which is unexpected Since my implementation used the simd, and if i invoke PyUnicode_DecodeLatin1 directly in PyUnicode_FromUCS1, it's slower too. @penguin-wwy do you have any ideas?

[penguin-wwy]

Could you describe the testing method? The tests I wrote myself do not have this issue.

# integration_tests/cpython_benchmark/fury_benchmark.py
STRING = "sjuveaibngurbzsivbrubiasb3r93284r92r1209130r0fa;2''j93r2nfln''[]\=-_+/,./!@$#%^&*()i9124u0hpq[jnzj0r9h034-2iu1058]"

def micro_benchmark():
    runner.bench_func(
        "fury_string", fury_object, language, not args.no_ref, STRING
    )
    runner.bench_func(
        "fury_large_string", fury_object, language, not args.no_ref, STRING * 10000
    )

Using PyUnicode_FromUCS1:
fury_string: Mean +- std dev: 54.7 us +- 2.5 us
fury_large_string: Mean +- std dev: 255 us +- 24 us

Using Fury_PyUnicode_FromUCS1:
fury_string: Mean +- std dev: 53.8 us +- 2.0 us
fury_large_string: Mean +- std dev: 236 us +- 6 us