[Docs] Add datatype × hardware support matrix #2512
Description
Is your feature request related to a problem? Please describe.
It's difficult to determine which datatypes and features are supported on which GPU architectures. Information is currently scattered across:
• High-level mentions in README (e.g., "FP8 on Hopper, Ada, and Blackwell")
• Inline code conditionals (e.g., get_device_compute_capability() >= (9, 0))
• Test skip conditions (e.g., pytest.skip("THD is only supported on Hopper+"))
• Error messages at runtime
Describe the solution you'd like
Add a clear Hardware Support Matrix page to the documentation with tables for "Datatype Support by Architecture" and "Feature Support by Architecture" (e.g. THD format).
Describe alternatives you've considered
Automatic capability detection at import - Could print a summary of available features, but doesn't address pre-deployment planning.
Additional context