Optimize fast-float exponential presentation path

[Neomelt]

Summary

• Optimize the default fast-float exponential presentation path by avoiding the extra write-and-swap step in the pointer fast path.
• Reuse write_significand directly in that path to reduce redundant memory operations.

Rationale

Issue: #4518

This change targets the post-Dragonbox presentation stage (not Dragonbox core generation), which is where #4518 suggests optimization opportunities.

Validation

• format-test: pass
• format-impl-test: pass
• quick perf sanity check shows positive direction locally (non-rigorous sanity benchmark):
  • before: ~355.1 ms
  • after: ~332.5 ms

Notes

• Small, behavior-preserving change.
• No formatting semantics intentionally changed.

[Neomelt]

I explored a second follow-up optimization in the same area and ran additional local microbenchmarks, but the gain was not stable across data patterns, so I intentionally did not include that change in this PR.

Keeping this PR focused on the small behavior-preserving improvement only.

If maintainers prefer a specific benchmark setup for #4518 (e.g. format-benchmark scenarios), I can run and post those numbers as a follow-up.