fix: use torch.amp via compatibility shim to avoid deprecated torch.cuda.amp warnings

[xyf5432]

Fixes #3517 — torch.cuda.amp.autocast has been deprecated since torch 2.4 and torch.cuda.amp.GradScaler since torch 2.3.

Summary

Both APIs still work but emit a FutureWarning when used; the
replacement device-agnostic APIs live in torch.amp.

Why migrate now, beyond warning noise: two entry points
(funasr/utils/export_utils.py, funasr/bin/realtime_ws.py) already
call warnings.filterwarnings("ignore") at module level, so the
FutureWarning is not visible on those paths today. The real risk is the
announced removal of torch.cuda.amp ("will be removed in a future
release", no version pinned; still present in 2.11): the moment torch
drops it, all 37 from torch.cuda.amp import ... sites in this repo
raise ImportError at import time — a hard failure no warning filter
can mask.

Changes:

• Adds funasr/utils/amp.py, a compatibility shim that re-exports
  torch.amp names when available (torch >= 2.3) and falls back to
  torch.cuda.amp otherwise.
• The shim supplies device_type="cuda" by default for autocast,
  matching the old torch.cuda.amp.autocast signature — existing call
  sites (with autocast(enabled=..., dtype=...)) work unchanged.
• Replaces all 37 from torch.cuda.amp import ... imports with the shim.
• Rewrites the 7 torch.cuda.amp.autocast(...) attribute call sites
  (llm_asr/model.py ×6, trainer_ds.py ×1) to use the imported name, so
  the deprecation path is actually avoided.

Type of change

• [√] Bug fix — deprecated API usage that becomes a hard ImportError when torch removes torch.cuda.amp
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

Verified on torch 2.11.0 with warnings.simplefilter("error", FutureWarning)
(re-asserted after importing funasr — export_utils.py / realtime_ws.py
set a catch-all filterwarnings("ignore") at module level that would
otherwise mask the test filter):

• from funasr.utils.amp import autocast, GradScaler — no FutureWarning

• with autocast(): / with autocast(enabled=True, dtype=None, cache_enabled=False): — no FutureWarning

• GradScaler(enabled=True) — no FutureWarning

• explicit device_type and positional args still pass through

• torch < 2.3 fallback path is identical to today (not yet deprecated there)

• [√] python -m compileall funasr examples tests — passes. (The SyntaxWarnings in funasr/models/fsmn_kws/encoder.py are pre-existing invalid escape sequences; that file is not touched by this PR.)

• Docs or links checked

• Runtime/deployment command tested

User impact

Users of FunASR on torch >= 2.3 — including the current PyPI default
torch 2.11, which the documented README install path resolves to — no
longer see the deprecation FutureWarning on training/inference paths,
and every entry point keeps importing when torch eventually removes
torch.cuda.amp. No behavior change on torch < 2.3.

Notes for reviewers

• The shim mirrors the old torch.cuda.amp.autocast signature via a
  device_type="cuda" default, because torch.amp.autocast requires a
  positional device_type.
• Migration boundary is 2.3, not 2.0: the shim imports both autocast
  and GradScaler from torch.amp together, and GradScaler only
  exists there since 2.3.
• bat/model.py keeps its torch < 1.6 guard — the shim's fallback
  imports torch.cuda.amp, which only exists since 1.6.
• The two entry points that already silence all warnings see no visible
  change; the fix is future-proofing for the removal.

Reference

• PyTorch AMP docs — official deprecation notice (torch.cuda.amp.autocast / GradScaler deprecated since 2.4 / 2.3, "will be removed in a future release")
• ultralytics/yolov5#13244 — same torch.cuda.amp → torch.amp migration, merged