v0.2.0

fexact is now written in Rust and built with maturin. Wherever the C
extension was correct, fexact returns the same p-values to the bit and the
same errors (checked on ~15k recorded cases), about twice as fast.

New

• fexact_many(arrays, workspace=200000, hybrid=None, threads=None): p-values
  of many tables, computed in parallel without the GIL. Raises
  ValueError("table i: ...") for the first table that fails.
• fexact_simulate(array, samples=10000, seed=None, threads=None): Monte Carlo
  p-value, like R's fisher.test(simulate.p.value = TRUE). Multithreaded,
  reproducible for a given seed whatever the number of threads, with no
  workspace limit.
• A Rust API (fisher_exact, fisher_exact_many, fisher_simulated, and
  compat::fexact with the C signature).

Fixed

• Wrong p-values on wide tables: the C code's 32-bit path counts overflowed
  silently, e.g. [[10, 6, 7] * 7, [5, 7, 9] * 7] gave 0.781 instead of
  0.874 (and 0.329 instead of 0.929 with 30 columns). They are now exact
  (checked against Monte Carlo simulation and R).
• Tables with large margins that failed with "FEXACT error 501" (hash key
  overflow), such as [[100000, 6, 7], [100000, 7, 9]], are now computed:
  node keys are 64-bit.
• Counts beyond 2^31 were silently wrapped to 32 bits (2**32 + 5 was read
  as 5). They are now read as 64-bit: fexact reports Out of workspace.
  for them, and fexact_simulate handles them.

Changed

• Some tables that raised ValueError now return p-values (see Fixed).
• The error "bug negative key" is now "integer overflow in exact computation"
  (R's message), raised only if a search key would overflow 64 bits.
• Building from source needs a Rust toolchain; wheels are provided for Linux
  (x86_64, aarch64), macOS (x86_64, arm64) and Windows (x64), for Python 3.9
  and later.
• The license metadata is GPL-3.0-only (unchanged terms).