feat: remove "alloc" feature requirement from base_convert/fmt

[DaniPopes]

Motivation

Alternative to #478 by making to_base_be not allocate directly.
Closes #478.

Needs #487 to gauge perf impact.

Closes #294.

Solution

PR Checklist

• Added Tests
• Added Documentation
• Updated the changelog

[codspeed-hq]

CodSpeed Performance Report

Merging #488 will not alter performance

_{Comparing DaniPopes:no-alloc-fmt-2 (d349967) with main (fb6460a)}

Summary

✅ 295 untouched benchmarks

[DaniPopes]

As mentioned in #490 this a lot slower because of precomputing the highest power, then dividing that power and (bigint) division is slow; it looks like just computing le and reversing iteration is almost always faster. Do we want to feature gate the slow implementation to not(feature = "alloc") and keep the collect one otherwise?

[prestwich]

As mentioned in #490 this a lot slower because of precomputing the highest power, then dividing that power and (bigint) division is slow; it looks like just computing le and reversing iteration is almost always faster. Do we want to feature gate the slow implementation to not(feature = "alloc") and keep the collect one otherwise?

Yes, i would appreciate this

[DaniPopes]

Done

[DaniPopes]

the performance improvement in fmt/4096 is due to hoisting the self.base load to outside of the loop which has drastic consequences to the generated code: https://godbolt.org/z/Wzbsbf4Pb

and the ones for 64 and 128 are due to the added fast paths

i'll move these changes to a separate PR