Stop auto-routing aarch64 3x3 convs to the indirect path

The indirect convolution round-trips the c_out accumulator through memory
once per (tap, in-channel), so it loses to the blocked-GEMM im2col path for
any non-trivial channel count. Gate it on output channels, off by default;
YSCV_INDIRECT_MAX_COUT re-enables it for narrow convs.

A53 yolo11n (4 threads): 3.0s -> 1.1s, tracker unchanged.

Author: enthropy7

crates/yscv-onnx/src/runner/conv/mod.rs

@@ -473,14 +473,24 @@ fn conv_compute_nhwc(
         // simple strided indexing.
         #[cfg(target_arch = "aarch64")]
         {
-            // The first-layer 3-channel 3×3 stride-2 Conv routes to its
-            // dedicated microkernel (via conv2d_nhwc_padded below), which holds
-            // the c_out accumulators in registers across the 27 taps. The
-            // generic indirect path round-trips the accumulator through memory
-            // once per (tap, in-channel) — fine when c_in is wide enough to
-            // amortise it, but ~27× off peak at c_in = 3.
+            // The indirect path round-trips the c_out accumulator through memory
+            // once per (tap, in-channel), so it loses to the blocked-GEMM im2col
+            // below once the channel counts are non-trivial. It is off by default;
+            // YSCV_INDIRECT_MAX_COUT sets an output-channel ceiling under which it
+            // is used. The first-layer 3-channel 3×3 stride-2 Conv keeps its own
+            // register-blocked microkernel in conv2d_nhwc_padded.
             let is_first_layer_3ch = input_nhwc.shape()[3] == 3 && sh == 2 && sw == 2;
-            if kh == 3 && kw == 3 && group == 1 && !cfg!(miri) && !is_first_layer_3ch {
+            let indirect_max_cout = std::env::var("YSCV_INDIRECT_MAX_COUT")
+                .ok()
+                .and_then(|v| v.parse::<usize>().ok())
+                .unwrap_or(0);
+            if kh == 3
+                && kw == 3
+                && group == 1
+                && !cfg!(miri)
+                && !is_first_layer_3ch
+                && o_ch <= indirect_max_cout
+            {
                 let t = yscv_kernels::conv2d_nhwc_indirect_padded(
                     input_nhwc, w_nhwc, bias, sh, sw, pt, pl, pb, pr, activation,
                 )