Do not inline collinear_chaining into get_chains

[ksahlin]

When collinear_chaining is inlined into get_chains, it seems the compiler
generates a slower DP loop. Marking it #[inline(never)] seems to make
chaining a tad faster.

Output is unchanged.

Speedup on an Apple M4 Max (rustc 1.97.0), -x, 1 thread, median of 3 runs:

• chrY: chaining -2.6% to -3.2%, total mapping time -1.6% to -2.2%
• CHM13: chaining -2.4% to -4.6%, total mapping time -0.5% to -1.2%
• maize (B73): chaining -2.8% to -4.3%, total mapping time +0.7% to -2.1%
• maize (Mo17): chaining -2.1% to -3.3%, total mapping time -0.6% to -0.8%
• rye (Lo7 and Lo7_V3): chaining -3.9% to -5.2%, total mapping time -1.3% to -2.3%
• fruit fly: chaining -3.1% to -4.0%, total mapping time -0.3% to -0.9%

The measurements are from one machine and compiler (Apple M4 Max, rustc 1.97.0),
so it may not hold across other environments.

[marcelm]

I’ve spent too much time benchmarking this change now, so I’ll just merge it because the diff is small. The difference is minute, but there seems to be one. I’m generally not so keen on overriding compiler decisions because the compiler usually knows better, and even if it helps now, it is possible that this optimization turns into a pessimization when we upgrade the compiler. But for now, it does improve things a tiny bit. We just need to re-benchmark this when we update the compiler.

[ksahlin]

Sounds good!