Add Q3_K_M weight dequant support

[unamedkr]

The GGUF weight loader (src/engine/tq_gguf_quants.c) handles Q2_K through Q6_K dequantization, but Q3_K_M (3-bit with medium grouping) needs testing and verification against the llama.cpp reference implementation.

What to do:

• Verify that the dequantize_q3_K path correctly handles the Q3_K_M nibble ordering and scale layout.
• Compare output against refs/llama.cpp/ggml-quants.c to ensure bit-level compatibility.
• Add a unit test in tests/ that roundtrips a known vector through Q3_K_M and checks MSE.

Files to touch: src/engine/tq_gguf_quants.c, tests/ (new test file or add to existing).