perf: optimize shift implementations

[DaniPopes]

Optimize shift left/right for better codegen:

• Replace the two-step carry trick (x >> (64 - bits - 1)) >> 1 with x.unbounded_shr(64 - bits) in the generic loop, exposing the funnel shift pattern to LLVM.
• Add specialized fast paths for u64, u128, and u256 in wrapping_shl/wrapping_shr via the as_primitives! macro.
• The u256 path decomposes into two u128 halves with branchless mask blending, producing shld/shrd + cmov with zero branches in the hot path.

llvm-mca (Zen 4) for U256::wrapping_shl:

• 57 instructions, 73 uOps, ~13.1 cycles latency, IPC 4.35
• Compared to native shl i256: 36 instructions, 37 uOps, ~15.1 cycles latency, IPC 2.39
• Our version has better latency due to avoiding store-forwarding penalties; native shl i256 has better throughput due to fewer instructions.

[codspeed-hq]

Merging this PR will degrade performance by 14.14%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 5 improved benchmarks
❌ 1 regressed benchmark
✅ 380 untouched benchmarks

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	most_significant_bits/4096/4096	29 µs	26.1 µs	+11.11%
❌	pow/128	64.9 µs	75.6 µs	-14.14%
⚡	to/f32/4096	29.1 µs	26.2 µs	+11.34%
⚡	to/f64/4096	29.2 µs	26.3 µs	+11.06%
⚡	wrapping_shl/128	281.9 µs	249.1 µs	+13.16%
⚡	wrapping_shr/128	292.9 µs	249 µs	+17.62%

---

_{Comparing DaniPopes:dani/better-shifts (441e078) with main (bff85c8)}

[DaniPopes]

only shr benches are impacted by this change, rest is noise: #573 (comment)