Feat: make TF32 and cuDNN benchmarking opt-in

[carshadi]

This PR updates CUDA backend configuration so performance-oriented PyTorch backend flags are controlled explicitly by InferenceConfig and default to False.

Previously, GpuWorker always enabled torch.backends.cudnn.benchmark, and use_tf32=True also forced both matmul TF32 and cuDNN TF32 on. This made backend behavior less explicit and could override PyTorch/session defaults. The new behavior sets only the intended backend options directly from config:

torch.backends.cuda.matmul.allow_tf32 = self.cfg.use_tf32
torch.backends.cudnn.benchmark = self.cfg.cudnn_benchmark

Changes

• Changed InferenceConfig.use_tf32 default from True to False
• Added InferenceConfig.cudnn_benchmark, defaulting to False
• Stopped modifying torch.backends.cudnn.allow_tf32, which defaults to True in PyTorch. This can be set to False by the user in their inference script if desired.
• Replaced CLI opt-out flag --no-tf32 with opt-in --tf32
• Added opt-in CLI flag --cudnn-benchmark
• Updated CLI docs to reflect the new flags
• Added config tests covering the new defaults and override behavior

Rationale

TF32 matmul and cuDNN benchmarking can improve throughput in certain cases, but it is hardware and data-size dependent. TF32 is only supported for Ampere and newer GPUs, whereas the default CodeOcean flex instance uses an older T4 (Turing). cuDNN benchmarking can also introduce a performance hit for smaller volumes where the auto-tuner overhead eclipses any performance gain.

[camilolaiton]

Looks good to me! Super useful! Probably not here in this repo, but we should discuss about quantization of our models (yours, Anna's, mine, etc) and if it requires any changes here (which I doubt).

[carshadi]

Looks good to me! Super useful! Probably not here in this repo, but we should discuss about quantization of our models (yours, Anna's, mine, etc) and if it requires any changes here (which I doubt).

Yeah I totally agree re: quantization