feat: support parser rounding modes

[tisonkun]

Summary

• Add a public RoundMode with five non-redundant modes: Ceil, Floor, HalfCeil, HalfFloor, and HalfEven.
• Add a non-exhaustive ParseOptions with a public round_mode field and HalfCeil as its default.
• Add ByteSize::parse_with(input, options); standard and Serde string parsing preserve the existing ties-to-larger behavior.
• Apply the unit multiplier first, then round exactly once to a whole byte without floating point.
• Keep the default parser free from runtime rounding state, track later discarded digits only for modes that need exact-tie detection, and scale long decimal fractions without per-digit division.
• Cover every mode with boundary, unit, overflow, and exact-arithmetic property tests; add isolated benchmarks for FromStr and parse_with.
• Document the contract in the API overview, README, and changelog.
• Verify with cargo x test, cargo x lint, and cargo x bench.

Design Notes

The rounding semantics are adapted from Jiff’s RoundMode, but the API does not copy all nine variants. ByteSize only accepts non-negative values, making Expand equivalent to Ceil, Trunc equivalent to Floor, HalfExpand equivalent to HalfCeil, and HalfTrunc equivalent to HalfFloor. The public enum keeps one direction-based name for each distinct behavior relevant to this domain.

The parsing API follows the cronexpr ParseOptions pattern: configurable behavior lives in a defaultable, non-exhaustive options type. FromStr uses the same HalfCeil contract through a zero-sized internal strategy, while parse_with uses the selected runtime mode. This preserves one parsing algorithm without carrying runtime mode state through the default integer scan.

Controlled same-machine A/B runs used matching benchmark binaries and at least one second per case. Against main, every existing FromStr case was at parity or faster: high-precision decimal parsing improved by about 17%, high-precision binary by about 1%, and the remaining cases by about 1–11%. Against the pre-optimization PR head, all five parse_with modes improved by about 8–14%.