perf: ARM CPU conv kernels for DocTR detection (rebased onto temp-9341)

[olyasir]

What

ARM CPU conv kernels for the DocTR DBNet detector, rebased from #152 (which
targeted temp-8828) onto temp-9341.

Speeds up DocTR detection when it runs on the ggml CPU backend — the hybrid
path used on Mali, where full-Vulkan detection is pathological. All kernel
changes are ARM-CPU only and bit-identical in their numerical formulas; GPU
backends and non-ARM CPU paths are untouched.

Changes vs temp-9341

ops.cpp / ops.h (conv kernels, applied unchanged from #152):

• NEON-vectorised conv_2d_dw_whcn (was fully scalar).
• Row-parallel im2col + a NEON-tiled, two-level-blocked 1×1 im2col transpose;
  falls back to the channel-strided split when rows_total < nthreads so
  1-D / OH==1 convs stay multi-threaded.
• Parallelised conv_transpose_2d and a contiguous-row UPSCALE fast path.
• ggml_compute_forward_conv_2d_fused / ggml_compute_forward_add_unary_fused:
  CONV_2D → ADD(bias) [→ UNARY] and standalone ADD(bias) → UNARY in one pass.

ggml-cpu.c (the one reconciliation):

• The CONV_2D → ADD → UNARY(relu|hardswish) epilogue fusion and standalone
  ADD → UNARY pair fusion now plug into temp-9341's ggml_cpu_try_fuse_ops()
  dispatcher (next to RMS_NORM+MUL), instead of perf: ARM CPU conv kernels for DocTR detection #152's inline block in
  ggml_graph_compute_thread.
• The F16-src1 mul_mat handling from perf: ARM CPU conv kernels for DocTR detection #152 is dropped — it only existed to
  keep mul_mat identical to perf: ARM CPU conv kernels for DocTR detection #152's older base, and is unreachable for OCR
  (f16/f16) and LLM (f32 activations). temp-9341's mul_mat is left as-is.

tests/test-backend-ops.cpp:

• CONV_2D_BIAS_ACT and ADD_BIAS_ACT cases covering both fusion paths
  (relu / hardswish, with and without the trailing activation). The reference
  CPU backend runs with use_ref=1 (fusion disabled), so these compare the
  fused result against the unfused reference.

Testing

• test-backend-ops test -b CPU: 15759/15759 OK (clang-22, x86 host),
  including the new fusion cases.
• NEON kernel paths are exercised by aarch64 CI (compiled out on x86).

Measured on a Pixel 9 Pro (Mali-G715), DocTR clinical_chemistry, warm:
detection 1.42s → ~0.83s on the CPU backend, output unchanged (boxes=197,
12/12 expected keywords).