[PPU][MTHREADS] Update low-precision float (fp8/fp4) support via backend capability contract

[zeroherolin]

Summary

This PR puts low-precision float (fp8/fp4) support on the PPU and MThreads backends under an explicit capability contract, and extends it where gaps remained. Each backend now declares the casts and dot combinations it actually supports; the semantic layer rejects undeclared combinations with the backend's diagnostic and warns at compile time whenever a non-native path runs — nothing degrades silently anymore. Pre-existing paths (PPU's FP16-promotion fp8 dot, MThreads' SQMMA/WMMA fp8 lowering with FMA fallback, fp8 casts, fp4 upcast) keep their behavior, now declared and fixed where broken; newly added are fp8e4nv on PPU cap80–88 via a bit-exact software cast, a working FP4→fp16 upcast below cap89, a rewritten interpreter fp-conversion baseline, and runnable fp8/fp4 tutorials.

Changes

• Common
  • new: DotCap/DotSupport (NATIVE / NON_NATIVE / UNSUPPORTED) with optional resolve_dot / resolve_dot_scaled codegen queries — UNSUPPORTED combinations raise with the backend's diagnostic, NON_NATIVE ones emit a compile-time warning (native=false).
  • new: capability declaration fields supported_fp8_cast_dtypes / custom_cast_fp8_dtypes (cast whitelist and software-cast routing) and async_copy_dtypes / descriptor_dtypes (data movement; declaration-only at this stage, shipped with h.metadata).
  • new: runnable tutorials python/tutorials/precision/01-fp8.py / 02-fp4.py covering capability queries, saturating casts, fp8 matmul, FP4 decode and dot_scaled.
  • unified: fp8 cast routing generalized from the hardcoded fp8e4b15-only special case to the declared whitelist.
  • fixed: interpreter float conversion rewritten with strict RTNE/RTZ and per-format fp8 special-value handling (NaN codes, fnuz, no-inf saturation).
• PPU
  • new: fp8e4nv opened on cap80–88 through a bit-exact software cast (frontend convert_custom_float8_sub89 for downcasts, inline-asm upcast core in the backend); native cvt from cap89 unchanged. The type was previously rejected outright below cap89.
  • fixed: FP4→fp16 upcast gated on capability — it unconditionally emitted cap89-only e4m3 cvt instructions; below cap89 it now decodes via the bf16 table with an exact integer rebias.
  • unified: the existing FP16-promotion fp8 dot is declared NON_NATIVE below cap89 and warns; resolve_dot_scaled declares e2m1 native on cap89 and NON_NATIVE decompose otherwise.
• MThreads
  • unified: fp8 storage/cast dual whitelist regularized (custom_fp8_dtypes → custom_cast_fp8_dtypes, cast whitelist derived from storage); the existing SQMMA/WMMA fp8 lowering is declared native for same-type fp8 with an fp32 accumulator on instruction-tile shapes from cap31, the existing FMA fallback is declared NON_NATIVE with a warning; dot_scaled always declares NON_NATIVE decompose; enable_fp8_burst2 moved from a C++-read env var into options; spec overlay kept in sync.
  • fixed: SQMMA operand retyping for fp8 staged through same-width bit containers; WMMA fp16-accumulator bail-out to the FMA fallback; WMMA same-element-type verifier.
• Tests: PPU fp8 compile-error tests updated for cap80 semantics (fp8e4nv compiles and warns non-native); exhaustive software-cast checks, including strict-RTNE downcast and clamping, enabled on cap80–88.

Compatibility

• The contract is optional: backends that do not register it keep the legacy dot/cast rules, and defaults preserve existing behavior for all other backends.
• Interpreter-mode numerics change for every backend as a correctness fix: strict RTNE (previously round-half-up) and per-format fp8 special-value handling.

Supersedes #1112 / #1113 / #1114.