CUDA extension loading should be modular

Describe the bug

TE's CUDA extension loading uses an all-or-nothing approach that causes the entire transformer_engine.pytorch module to become broken when any optional dep (e.g. flash-attn) fails.

Root cause: Version checking and optional dependency imports happen during core CUDA extension loading. When an optional dependency check fails (e.g., flash-attn 2.8.3 > maximum supported 2.8.1), the exception is caught silently, leaving te.pytorch partially initialized.

Steps/Code to reproduce bug

# 1. Install TransformerEngine 2.9.0 and flash-attn 2.8.3
pip install transformer-engine==2.9.0
pip install flash-attn==2.8.3  # Note: TE 2.9.0 only supports <= 2.8.1

# 2. Try to use TransformerEngine
python3 -c "
import transformer_engine as te
print(f'TransformerEngine imported: {te.__version__}')
print(f'Has pytorch attr: {hasattr(te, \"pytorch\")}')

# Try to use functionality that doesn't need flash-attn
try:
    linear = te.pytorch.Linear(128, 256)
    print('✓ te.pytorch.Linear works')
except AttributeError as e:
    print(f'✗ Error: {e}')

Output:
See https://github.com/NVIDIA/Megatron-LM/issues/696#issuecomment-3558275306

Expected behavior

1. Modular loading: Core extensions should load independently of optional extensions with feature flags (e.g. HAS_FA3).
2. Warnings instead of silent failures and hard version pinning

# If flash-attn version incompatible
UserWarning: flash-attn 2.8.3 is newer than tested maximum 2.8.1. This may work, but report issues if encountered.

Environment overview (please complete the following information)

Python: 3.10.12
Transformer Engine: 2.9.0
flash-attn: 2.8.3

Start by reproducing the documented import with Transformer Engine 2.9.0 and flash-attn 2.8.3, then trace transformer_engine.pytorch initialization and its CUDA extension and optional-dependency checks. Done means core functionality such as te.pytorch.Linear remains available when flash-attn is incompatible, while optional features expose flags and emit warnings instead of failing silently.