Phase 1 — Core ggml type system:
    - Add GGML_TYPE_Q4_T = 43 enum in ggml.h, bump COUNT to 44
    - Define QK_T=256 and block_q4_t struct (d[FP16] + qs[128 nibbles], 130 byte
s)
    - Register type_traits[] entry with dequantize_row_q4_t / quantize_row_q4_t_
ref
    - Implement dequantize_row_q4_t(): same formula as Q4_0, block size 256
    - Implement quantize_row_q4_t_ref() and quantize_row_q4_t() scalar CPU impl
    - Add validation case in ggml_validate_row_data() for type safety

    Phase 2 — CPU backend:
    - Add type_traits_cpu[] entry with vec_dot = ggml_vec_dot_q4_t_q8_0
    - Implement ggml_vec_dot_q4_t_q8_0_generic() kernel (handles QK_T=256 blocks
      mapped to Q8_0 keys, fixing NaN in dot product test)

    Phase 3 — CLI wiring:
    - Add LLAMA_FTYPE_MOSTLY_Q4_T enum in llama.h (value 41)
    - Wire ftype->ggml_type mapping in llama-quant.cpp (Q4_T -> GGML_TYPE_Q4_T)
    - Register Q4_T string option in quantize.cpp QUANT_OPTIONS array

    Phase 4 — End-to-end validation:
    - Add tensor fallback: Q4_T -> Q4_0 for tensors with ncols not divisible by
256
    - Full pipeline verified: f32->Q4_T quantization, GGUF load, and inference t
est

    Spec: flat block layout, same dequant as Q4_0 ((nibble-8)*d), ~4.06 bpw,
    QK_T=256 blocks. Compatible with existing Q8_K matmul via vec_dot dispatch.

    Tests: test-backend-ops OK, test-gguf 71/71 OK, test-quantize-fns Q4_T OK