[multifn] Optimize non-cached effective-dispatch-value

[alexander-yakushev]

Further improvements to method dispatching machinery. When it is cached, this shouldn't matter, but I keep encountering cases where it is not cached and thus it matters a bit.

For the benchmark, I've accumulated a sample of ~20k argument sets to effective-dispatch-value I got these results:

Before:
Time per call: 13.83 ms   Alloc per call: 31,446,785b
Time per call: 13.82 ms   Alloc per call: 31,446,785b

After:
Time per call: 7.44 ms   Alloc per call: 8,638,718b
Time per call: 7.40 ms   Alloc per call: 8,638,718b

Normalized to a single invocation, this means that a single invocation allocations dropped from 1.39KB to 380B which looks significant enough for something as basic as a multimethod invocation.