Logits processors: Update inplace, with batch operation

[lapp0]

Changes

For GuideLogitsProcessor,

• Update logits inplace: [7514aff]
  • Fixes Update logits array in-place #859
• Batch update logits: [8aa0b0d]
  • (Because I noticed GPU profile was worse than CPU due to cuda synchronization)

Benchmarks for both changes

CI doesn't benchmark torch_cuda, so I've included it here.

Update logits inplace

Before [245c7fc]	After [7514aff]	Ratio	Benchmark (Parameter)
159_0.4μs	181_1μs	1.14	time_structured_generation('numpy', 'Z*')
149_0.3μs	170_0.7μs	1.14	time_structured_generation('torch', 'Z*')
292_0.8μs	254_1μs	0.87	time_structured_generation('torch_cuda', 'Z*')
572_2μs	391_1μs	0.68	time_structured_generation('torch_cuda', '[^Z]*')

Batch update logits

Before [245c7fc]	After [8aa0b0d]	Ratio	Benchmark (Parameter)
481_5μs	401_2μs	0.83	time_structured_generation('numpy', '[^Z]*')
466_3μs	386_1μs	0.83	time_structured_generation('torch', '[^Z]*')
159_0.8μs	106_0.5μs	0.67	time_structured_generation('numpy', 'Z*')
149_0.7μs	94.7_0.2μs	0.64	time_structured_generation('torch', 'Z*')
290_1μs	149_0.4μs	0.51	time_structured_generation('torch_cuda', 'Z*')
573_3μs	229_1μs	0.4	time_structured_generation('torch_cuda', '[^Z]*')

Testing

• Tested on CUDA machine: All tests pass
  • (Except for failures on main due to a recent vLLM change disallowing opt125)
• Did not test on Metal.