feat: honour ACESTEP_DTYPE env var on standard CUDA devices

[Copilot]

ACESTEP_DTYPE was read nowhere — the env var existed but had no effect on standard (non-ROCm) CUDA paths, leaving users with no way to override the hardware-inferred dtype.

Changes

init_service_orchestrator.py

• In the non-ROCm elif resolved_device == "cuda" branch, check ACESTEP_DTYPE before hardware auto-detection.
• Accepts float32 | float16 | bfloat16; invalid/unset values fall through to the existing Ampere/Pre-Ampere detection unchanged.

env_dtype_str = os.environ.get("ACESTEP_DTYPE", "").strip().lower()
if env_dtype_str in ("float32", "float16", "bfloat16"):
    self.dtype = getattr(torch, env_dtype_str)
    logger.info(f"[initialize_service] ACESTEP_DTYPE={env_dtype_str} override: using dtype={self.dtype}.")
elif gpu_config.cuda_supports_bfloat16():
    self.dtype = torch.bfloat16
else:
    self.dtype = torch.float16
    ...

init_service_loader.py

• Pre-Ampere CUDA branch in _load_main_model_from_checkpoint now checks self.dtype before selecting attention implementation.
• float32 → sdpa (no overflow risk in SDPA's fused softmax with float32).
• float16 (default on Pre-Ampere) → eager workaround retained unchanged.

elif device == "cuda" and not gpu_config.cuda_supports_bfloat16():
    if getattr(self, "dtype", None) == torch.float32:
        attn_implementation = "sdpa"   # float32 is safe with SDPA
    else:
        attn_implementation = "eager"  # float16 overflow guard

Tests

• 7 new unit tests in init_service_test.py covering: all three valid ACESTEP_DTYPE values, invalid/unset fallback, and the float32 vs float16 attention-selection paths on Pre-Ampere CUDA.