nx_blockfloat_mha_gate.nx

nx_blockfloat_mha_gate.nx -- BLOCK-FLOAT weights through REAL MULTI-HEAD attention. Composes the existing gradcheck-verified multi-head forward (mha_fwd/mha_out from nx_nofloat_multihead_gate: RMSNorm -> Q/K/V -> split H heads -> per-head RoPE + scaled causal-softmax + A.V -> concat -> out-proj -> residual) UNCHANGED, run with the weight matrices (Wq,Wk,Wv,Wo) BLOCK-FLOAT quantized. Extends the single-head block result to multi-head -- the defining transformer feature -- proving block-float runs real multi-head attention better than per-tensor INT8. 1 multi-head attention RUNS with block-float weights (non-trivial output) 2 EXCEED: block-float MHA error vs full-precision < per-tensor INT8 3 DETERMINISTIC: block-float MHA run twice == bit-identical 4 per-tensor is LOSSY (output != full-precision) -- block-float's win is over real info loss expect_exit: 0 license_tier: ORIGINAL