bifrost 0.2.2

This update makes search initialization explicit and corrects progress-related
reproducibility and numerical-thread handling. Requires R 4.2 or newer.

• New initialization default: start_strategy = "legacy_1.2.1" calculates
  fresh, data-dependent BM/BMM starts using the mvMORPH 1.2.1 development
  strategy used in earlier bifrost analyses, including its random second-tip
  safeguard for singleton regimes. It supplies starts to the installed
  mvMORPH without pinning or modifying that dependency. This is development
  behavior, not exactly the CRAN 1.2.1 initializer.
• Native option: use start_strategy = "native" for mvMORPH's own
  initialization, including EmpBayes. Explicit start and grid.search = FALSE
  take precedence. Fits record the initialization strategy and mvMORPH version.
• Progress and threads: enabling or disabling progress preserves the same
  fitting random sequence for matched seeds and settings. Serial fitting
  preserves numerical-thread settings; concurrent fits request one numerical
  thread per worker. Console progress and IC-history displays are refined.

Results can differ from earlier releases. Reproducing a sequence of searches
requires the same random-state sequence, not merely resetting the same seed
for each independent run. The historical initializer retains scale and RNG
dependence; it does not guarantee the best fit. Some intermediate fits can
still fail to converge (#242).

The existing Gumbel KL-divergence bootstrap can discard valid large draws
through overflow/underflow and understate uncertainty. Do not rely on the
affected confidence interval as a dependable uncertainty estimate; its fix
is deferred (#231).

See the changelog and
release-status page
for details. Version 0.2.2 is also available on CRAN.