perf: hoist tile/clamp out of highp sampler inner loop

[bhark]

This PR is part of a series linked to #174.

What

Hoists tile() / clamp / y*stride out of the inner loop of the bicubic and bilinear samplers in highp. Dead sample() helper is removed. The two sampler functions have been collapsed, as they're very similar after this work.

Why

tile() is opaque to LLVM, so it never gets CSE'd out of the inner loop. Hoisting drops the cost per-stage, such that the inner body collapses to one i32 add + bitcast + gather + load + mad.

Results

	speedup
Geomean (56 benches)	1.01x
patterns::hq (bicubic, sampler N=4)	1.78x
patterns::lq (bilinear, sampler N=2)	1.32x
patterns::plain	1.00x
gradients::*, fill::*, blend::*	1.00x

No regressions. Benched on an i5-13400F with -Ctarget-cpu=haswell.

Notes

• The primitives (tile, ulp_sub, to_u32x8_bitcast) are unchanged. This is just a scheduling change.
• Should be bit-identical to the original sample() path.

[RazrFalcon]

Nice! There were no const generics in Rust when this code was originally written.