Three optimisations - two (very) small/minor, one medium, no public API changes:

Optimisations

1. Replaced s.chunks(4) with s.chunks_exact(4) for the most micro of micro-optimisations.

2. As per @tafia's review suggestion perf: cache path and attr lookups, reuse buffers, minimise allocs #606 (review), replaced use of the standard HashMap with hashbrown::HashMap (which uses ahash). This was already a transitive dependency via zip, so doesn't actually expand the current set of deps (though is now explicit in Cargo.toml). Also replaced BTreeMap with HashMap in two places where the map does not actually need to maintain ordering ("read_relationships" and "number_formats") to get O(n) instead of O(log n); marginal impact though as these maps are rarely (if ever?) going to be large.

Update: reverted explicit hashbrown integration (benefit was too marginal to reliably measure - maps are likely too small in relation to everything else), but kept the switch from BTreeMap to HashMap in places where order isn't needed..

3. Extended the same buffer-reuse optimisation that was part of perf: cache path and attr lookups, reuse buffers, minimise allocs #606 to XlsxCellReader, but cleaned it up a bit by consolidating the buffers in a new ValueBufs struct. No more use of read_string, so dropped it (only the variant that takes shared buffers is used now).

Miscellaneous

clippy complained about one of the functions having one too many args, so I also consolidated workbook-related args into a WorkbookContext struct; seems tidy?

Results

~6-7% speedup¹ when testing on a large (10 million cell) mixed-dtype Workbook.