refactor: split fp8 gemm device code into layered headers

- split gemm.cuh into gemm/{policy,load,scheduler,mainloop,epilogue}.cuh (humming/CUTLASS-style layering, files 28-336 lines); the umbrella keeps the kernel orchestrator, host planning and the gemm<> entry so ops.cu and the C tests build unchanged
- move the measured design essays (swizzle derivation, ring-depth barrier invariant, launch crossovers, NN swap) into an FP8 design-notes section in docs/developer/cuda_kernels.md, leaving one-line constraints at each symbol
- refresh the doc's FP8 file table and layout tree (fix stale mm.cu / fp8_mma_test.cu names)

- structure-only change: extension rebuilds identical, C tests all pass, tests/extension 65 passed, quantize layouts byte-exact, NT routing torch.equal, e2e M=8192 530.6ms / 1.26x unchanged

csrc/kernels/fp8/gemm/scheduler.cuh

@@ -0,0 +1,28 @@
+#pragma once
+// Tile scheduler: the linear CTA id maps to (block_m, block_n) in grouped
+// (L2-friendly) raster — consecutive CTAs share one B column stripe — or
+// plain N-fastest raster (kRasterGroup=0, the measured best for dX's
+// crosswise-B layouts where grouping was neutral).
+
+namespace astrai {
+namespace fp8 {
+
+template <int kRasterGroup>
+struct Fp8GemmTileScheduler {
+    static __device__ int2 tile(const uint3& block, const dim3& blocks) {
+        if constexpr (kRasterGroup > 0) {
+            constexpr int kGroupM = kRasterGroup;
+            const int bid = int(block.y) * int(blocks.x) + int(block.x);
+            const int group_first_m = (bid / (kGroupM * int(blocks.x))) * kGroupM;
+            const int group_rows =
+                min(int(blocks.y) - group_first_m, kGroupM);  // M-tail group is short
+            return int2{group_first_m + bid % group_rows,
+                        (bid % (kGroupM * int(blocks.x))) / group_rows};
+        } else {
+            return int2{int(block.y), int(block.x)};
+        }
+    }
+};
+
+}  // namespace fp8
+}  // namespace astrai