fix(spacemit): fix MiniCPM-V SMT multimodal inference

[oscar1229]

Fixes three independent bugs that prevented MiniCPM-V from running via the SMT media backend with multi-threaded warmup and multi-turn image conversations.

• fix(spacemit): fix IME paired-lane GEMM threadpool deadlock

  The IME GEMM kernels (forward_mul_mat and the mul_mat_id MoE path)
  rendezvous thread pairs (2k, 2k+1) on a spine_barrier built for two
  participants, so both lanes must call spine_barrier_wait() the same
  number of times. The old per-thread loop could iterate a different
  number of times per lane when gemm_n was not a multiple of
  NB_COLS*nth, and the trailing even thread on odd nth had no partner,
  so warmup hung with -t 8. Drive the loop from a pair-aligned base
  with a per-lane offset (both lanes always iterate equally; an
  out-of-range lane skips the GEMM but still hits the barrier) and
  guard the barrier with has_pair so a partnerless thread never waits.

• server: force full re-prefill for multimodal FULL-only KV cache reuse

  MiniCPM-V runs on the qwen35 hybrid (SSM + periodic full-attention)
  backend whose KV memory only supports full sequence removal. On a
  multi-turn request, partial prompt-cache reuse would either restore a
  context checkpoint (resurrecting a KV state inconsistent with the
  external smt/ONNX vision embeddings) or call partial memory_seq_rm on
  FULL-only memory, which returns false and triggers GGML_ABORT. When
  the context is multimodal and the reused prefix is partial, force a
  full re-prefill (pos_next = 0, n_past = 0) before the checkpoint /
  seq_rm path. Pure-append turns and non-multimodal contexts are
  unaffected.

• feat(mtmd): add MiniCPM-V SMT vision preprocessing

  The MiniCPM-V SMT vision ONNX export does not normalize pixels
  internally. Detect minicpmv / minicpm_v / minicpm-v architectures and
  route them through rgb_u8_to_chw_f32_with_config, which reads
  rescale_factor / image_mean / image_std from config.json's
  vision_preprocess block and emits a CHW float32 tensor. Target
  defaults to 448x448, overridable via vision_model.input_width/height.

Overview

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: