Backport v2’s retain implementation to v1

[charliermarsh]

Summary

This PR applies the existing v2 compaction approach to v1's retain: leave the initial retained prefix in place, then move each survivor into the next hole instead of swapping it with a discarded element. Removed elements are dropped during the scan, and a guard restores the untouched tail if the predicate or a destructor panics.

I benchmarked this on Linux (x86), and retaining half the input took 40-60% less time for small integer vectors, 16-21% less time for 4,096 integers, and 24-38% less for 4,096 larger elements:

Workload	Element size	Before	After	Paired time change
Keep half, 16 elements	8 B	39.33 ns	22.10 ns	−45.5% to −42.0%
Keep half, 17 elements	8 B	46.95 ns	20.97 ns	−59.9% to −54.1%
Keep half, 4,096 elements	8 B	17.205 µs	14.247 µs	−21.0% to −15.9%
Keep half, 4,096 elements	128 B	31.931 µs	21.497 µs	−37.8% to −23.6%
Keep all, 16 elements	8 B	39.79 ns	12.46 ns	−69.8% to −67.6%
Keep all, 17 elements	8 B	38.71 ns	9.93 ns	−80.3% to −73.0%
Keep all, 4,096 elements	128 B	19.722 µs	14.650 µs	−29.7% to −7.3%
Keep none, 16 elements	8 B	40.60 ns	17.24 ns	−57.6% to −55.8%
Keep none, 17 elements	8 B	40.94 ns	18.55 ns	−55.1% to −54.4%
Keep none, 4,096 elements	128 B	18.427 µs	17.594 µs	−7.0% to +13.1%
Unchanged Vec control: keep half, 4,096 elements	128 B	21.275 µs	21.486 µs	−9.6% to +10.8%

(I included some native benchmarks but removed in a follow-up commit; you can see them in the PR history.)

[alejandro-vaz]

can you paste the table with performance changes??

just to have it explicit on the PR

[charliermarsh]

Done!

[alejandro-vaz]

thanks @charliermarsh