Clamp the patch merge to ±2π in the native Swin encoders

NVIDIA's enc1 and enc2 clamp the patch merge's result to ±2π (as f16) before
its e4m3 encoding (max.f16x2 / min.f16x2 with 0fC0C90FDB / 0f40C90FDB after the
merge's last MMA). The native layers did not, in either template or set:
merge_clamp (swin_common.h) now does, at the three sites that encode the merged
tokens (swin_block.h's e4m3-activation and f16 paths, swin_tok.h), and
swin_check.py's model clamps the same way.

Measured on an RX 9070 XT against a run with no native kernels (preset L,
accuracy set, the flags of Ghost of Tsushima, 1280x720 -> 3840x2160, one
frame): native enc1 alone 34 dB and enc2 alone 31 dB from the translated
kernels before, byte-identical after; the whole native set 28 -> 80 dB (the
rest is dec1, 0.09% of the output values). Preset M's fast set against
translated M: 30.8 -> 35.1 dB after 30 frames. The gfx11 path has the same
change and was not run on hardware.

The repo's tests pass (36, 3 skipped).

kernels/tools/swin_check.py

@@ -47,7 +47,9 @@ def model_block(P, layer, bx, by):
             Wg = me.E4M3[P.wbuf[sm.woff_table(PM0 + g * 128 * C, 512, 64 * C, 4 * C, 32)]]
             bias = P.f16vec(PM0 + 4 * C * NPM + 64 * g, 32)
             cols.append(me.mma_chain(A, Wg, C0=np.broadcast_to(bias, (16, 32)).copy()))
-        merged = me.q8(sm.pair_to_nat(np.concatenate(cols, axis=1)))
+        # the merged tokens are clamped to +-2 pi (f16) before the e4m3 encoding
+        lim = float(np.float16(6.2831855))
+        merged = me.q8(np.clip(sm.pair_to_nat(np.concatenate(cols, axis=1)), -lim, lim))
     return out, merged