…esign

remove legacy GPTQ/clip optimization code and quantize_oq() full-model
path. only streaming quantization path remains.

fix dense model budget plan bug: gate_proj/up_proj excluded from
sensitivity boost due to dead MLP asymmetry reference.

…plementation

old turboquant had MSE-only codec with C++ metal extensions and was 350x slower
than SDPA. now imports from mlx-vlm which has multi-codec support (MSE, Prod,
Polar, Split), fractional bits (3.5), and 3 optimized decode paths via inline
mx.fast.metal_kernel. only BatchTurboQuantKVCache is implemented locally for
omlx's continuous batching scheduler. re-enables turboquant in admin UI.

omlx was missing ImageOps.exif_transpose() that mlx-vlm applies when
loading images. phone photos with EXIF rotation tags were passed to the
vision model in the wrong orientation, causing inaccurate recognition.

quantize KV cache immediately during prefill (not deferred to first decode
token) so peak memory is reduced throughout the entire inference. use
concat-based state growth instead of in-place write. eval logits instead
of cache states to work around mx.eval NamedTuple bulk traversal hang.
disable prefill_attention Metal kernels for D>128 (hangs on Qwen3.5-27B
with head_dim=256), fall back to dequantize+SDPA.

Qwen3.5-27B-4bit pp32768/tg128:
  baseline: peak 21.82GB, TTFT 85.0s, tg 30.6 tok/s
  TQ 3-bit: peak 21.11GB, TTFT 92.8s, tg 16.3 tok/s

batch-quantize every 32 decode tokens instead of per-token quantization
to improve GPU utilization (rotation matmul on 32 tokens vs 1 token).
decode attention uses hybrid approach: TQ Metal kernels for quantized
old tokens + standard dot product for buffered fp16 recent tokens.

also fixes prefill_attention: disable for chunked prefill because
mlx-vlm's value kernel unrolls n_repeats*L at compile time, hanging
the Metal shader compiler for large L (e.g. 2048).

Qwen3.5-27B-4bit pp32768/tg128 (3-bit TQ):
  before: TTFT 92.8s, tg 16.3 tok/s, peak 21.11GB
  after:  TTFT 92.8s, tg 19.2 tok/s, peak 21.11GB
  (baseline without TQ: tg 30.6 tok/s, peak 21.82GB)