fix: correct KV cache memory stats for K/V metadata and fp16 baseline

[dipeshbabu]

Fix KVCacheCompressor.memory_stats() so it matches the actual storage layout used by compress().

The previous implementation undercounted memory in two ways:

• it treated the original fp16 baseline as a single tensor instead of combined K+V
• it omitted stored norms from the compressed accounting, including the V-side norm and one of the K-side norms

What changed

• count original KV cache size as fp16 K + fp16 V
• count K compressed storage as:
  • d * k_bits
  • vector_norm
  • residual_norm
• count V compressed storage as:
  • d * v_bits
  • vector_norm

Tests

Added an exact regression test in tests/test_kv_cache.py that checks the byte math for:

• head_dim=128
• k_bits=3
• v_bits=3
• seq_len=1
• num_layers=1
• num_heads=1

Also verified the full KV cache test file passes.

[TheTom]

hey there. thank you for the contribution. i'll be getting to them. I apologize for the delay.