Add OSCAR 2-bit KV cache (PER_GROUP) support to CPU GroupQueryAttention

[Hector Li (HectorSVC)]

Description

Summary
This PR adds an opt-in 2-bit KV cache path to the GroupQueryAttention (GQA) contrib op on the CPU EP, implementing the OSCAR scheme (per-group asymmetric INT2 codec + a sink/recent high-precision window + optional spectral rotation). It extends the existing quantized-KV-cache machinery (which already supports INT8/INT4) rather than adding a new op, so existing GQA graphs are unaffected. All new behavior is gated behind new attributes / session-config entries and defaults to off.

Reference

OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization. 28 May, 2026
https://arxiv.org/abs/2605.17757

Motivation and Context

For long-context / large-batch serving, the KV cache dominates memory. INT8→INT4 halves it; going to 2 bits needs outlier handling. OSCAR keeps a small sink+recent window in full precision, quantizes the long tail of history to 2 bits with per-group scale/zero, and (optionally) applies an orthogonal rotation so the history quantizes with much lower error while leaving QK scores invariant.

What's included
New quantization path (k_quant_type/v_quant_type = "PER_GROUP", kv_cache_bit_width = 2)

Per-token, per-group asymmetric INT2 codec (4 codes/byte) with per-group scale + zero-point stored inline in the packed cache row (head_size/4 + num_groups2meta_bytes). Scales are computed dynamically at append time, so no k_scale/v_scale inputs are required.
kv_quant_group_size attribute selects the group size.
k_quant_rho / v_quant_rho: magnitude-percentile outlier clip before computing per-group min/max (1.0 = disabled).
kv_quant_metadata_fp16: store inline scale/zero as fp16 instead of fp32, shrinking the row (e.g. 48B→40B for D=128,G=64, i.e. 3.0→2.5 effective bits/elem) with no measured accuracy loss.
Mixed-precision sink/recent window (OSCAR "Option C")

New optional I/O: inputs past_hp_key/past_hp_value (16, 17), outputs present_hp_key/present_hp_value (4, 5) — the first sink and last recent tokens are kept unquantized (type T), the middle history is 2-bit.
New session-config entries gqa.kv_quant.sink / gqa.kv_quant.recent select the window sizes.
Optional spectral rotation

New optional inputs oscar_rotation_k/oscar_rotation_v (18, 19), per-kv-head orthogonal matrices (kv_num_heads, head_size, head_size) applied to post-RoPE K/Q (and V) before 2-bit quantization; the V output is un-rotated by the transpose.
fp16 compute support

MLFloat16 is registered for GQA; the OSCAR path bridges fp16 Q/K/V (and the hp window) to the float codec and converts outputs back, so an fp16 model can drive the 2-bit cache. (attention_bias and output_qk are not supported on the fp16 2-bit path and return a clear error.)

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.