optimize: use inline hint instead of inline(always) for enum codecs

[farhan-syah]

Fixes #60.

Changes the two derive-generated codec functions from #[inline(always)] to #[inline]. Every hand-written #[inline(always)] in zerompk/src/impl.rs and zerompk/src/lib.rs is left exactly as it is — those bodies are a few instructions on a hot path and the attribute is right for them.

These two are the only sites whose body size is chosen by the user's type rather than by the library. A derived enum decoder grows with the variant count, and forcing it inline means a parent enum that wraps several op enums becomes one
function containing all of their decoders. LLVM's passes are superlinear in function size, so the cost grows far faster than the type does — on the 50-line reproduction in #60, a crate-only rebuild goes from 28.90s to 0.82s.

#[inline] still lets the cost model inline the small cases, which is nearly all of them.

Verification

• zerompk's own suite: unchanged. The 5 failures in tests/derive.rs are present identically on a pristine main worktree, so they are not from this change.
• Downstream: a database codebase using zerompk for its internal wire format has a 6k-line crate of plan enums that went from 32 minutes to 44 seconds to compile, with its full suite of 13,633 tests — all of which exercise the encode/decode paths — passing unchanged.

I did not measure a runtime effect. The change only removes a forcing attribute, so small codecs are still inlined by the cost model.

[nuskey8]

Thank you for the PR.

I would like to accept this change, but the comment in the relevant section seems redundant. Could you limit the change to just the #[inline(always)] -> #[inline]?

[farhan-syah]

Thank you for the PR.

I would like to accept this change, but the comment in the relevant section seems redundant. Could you limit the change to just the #[inline(always)] -> #[inline]?

Done

[nuskey8]

Merged the PR. Thanks!