HGQ v2 support (part 1)

[calad0i]

Description

Preliminary support for the HGQ v2 package. Adds corresponding handlers, bit-exact optimizer, and layer implementations.

Includes the following:

• Keras v3 support ✅
  • Object based parser (and supports reused layers) ✅
  • Layer handlers for incompatible layers ⚠️
  • Make QKeras optional ✅
    • Lazy converter import ✅
• MultiHeadAttention (✅ parallel, stream ❌)
  • EinsumDense (parallel ✅, stream ❌)
    • General transpose (parallel ✅ stream ⚠️)
  • Einsum ✅
  • Bit-exact softmax ✅
  • Multi-dimentional softmax (parallel ✅ stream ❌)
  • (?) cossim based attn
• Automatic precision config
  • accum_t precision inference ✅
    • Symbolic precision converter (fixed ✅ minifloat ❌)
      • const_arr/precision_arr +-*@ precision_arr -> precision_arr
    • Support for specific layers (everything in HGQ v1 + alpha) ✅
  • Fuse RND into bias and cast to TRN when possible
• minifloat support ❌
  • minifloat * minifloat -> fixed_point cpp template
    • Check if the compiler is smart enough
    • Precision conversion rules

Type of change

• New frontend and features

Tests

N/A at the moment, not possible w/o test env update.
Bit-exact logic supersedes HGQ proxy embedding configs, thus those tests are reused w/o need of modification.

Checklist

• I have read the guidelines for contributing.
• I have commented my code, particularly in hard-to-understand areas.
• I have made corresponding changes to the documentation.
• My changes generate no new warnings.
• I have installed and run pre-commit on the files I edited or added.
• I have added tests that prove my fix is effective or that my feature works.