Replace deprecated torch.cuda.amp API with torch.amp equivalents

[xyf5432]

Fixes #2024

Summary

torch.cuda.amp has been deprecated since torch 2.3 (GradScaler) / 2.4 (autocast, custom_fwd, custom_bwd) and is scheduled for removal. This PR migrates the remaining usages to torch.amp, which has provided the same names since torch 2.3, and enforces the existing torch>=2.6.0 floor from requirements.txt inside setup.py.

Changes:

• apex/_autocast_utils.py — torch.cuda.amp.autocast_mode._cast(args, dtype) → torch.amp.autocast_mode._cast(args, "cuda", dtype). The torch.amp variant takes an explicit device_type since torch 2.5; the old two-arg form only forwards and warns "Please use torch.amp.autocast_mode._cast(value, 'cuda', dtype) instead".
• apex/contrib/cudnn_gbn/batch_norm.py — from torch.amp import custom_bwd, custom_fwd; decorators now pass the required keyword-only device_type="cuda" (same style as the existing apex/contrib/conv_bias_relu/conv_bias_relu.py).
• apex/contrib/optimizers/distributed_fused_adam.py — module-level aliases GradScaler = torch.amp.GradScaler and GradScalerOptState = torch.amp.grad_scaler.OptState (verified to be the identical objects behind the deprecated names), used by the grad_scaler type annotations, docstrings, and the per-method OptState assignments that are now removed.
• apex/contrib/test/optimizers/test_distributed_fused_lamb.py — from torch.amp import GradScaler.
• tests/L0/run_mlp/test_mlp.py — torch.amp.autocast("cuda", enabled=...).
• setup.py — compile-time check now rejects torch < 2.6.0 (via packaging.parse, previously a dormant 0.4 check) and install_requires gains torch>=2.6.0, matching requirements.txt.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Validation

Verified on torch 2.11.0 (CPU build):

• torch.amp.grad_scaler.OptState is torch.cuda.amp.grad_scaler.OptState → True, so the is comparisons in unscale_grads/step are unaffected.
• New _cast path is warning-free under warnings.simplefilter("error", FutureWarning), while the old two-arg _cast raises under the same filter — the migration is effective and the filter is sensitive.
• New and old _cast produce identical outputs for identical inputs.
• custom_fwd(device_type="cuda") / custom_bwd(device_type="cuda") forward + backward run cleanly.
• Module-level aliases resolve to the torch.amp objects; the GradScaler | None annotations evaluate.
• setup.py version check passes 2.11.0 and rejects 1.13.1 (control group raises RuntimeError).
• ruff format --check with the pinned 0.16.1: all touched files unchanged.
• ruff check: identical error set before and after (27 → 27, only line-number shifts from the added module-level aliases) — no new lint issues.
• py_compile passes on all touched files.

User impact

No API or behavior changes for users; the FutureWarnings emitted through torch.cuda.amp (and the risk of breaking on its removal) are eliminated.

Notes for reviewers

• torch.amp has provided the same names since torch 2.3, and requirements.txt already declares torch>=2.6.0, so the migration introduces no minimum-version regression.
• The three-argument _cast(value, device_type, dtype) signature exists since torch 2.5; 2.3/2.4 are outside the declared floor.