v1.16.0 — feat: two-phase backtest API, standardised fit() return contract

What's changed

New features

• BaseModel.wrap_fit_result(params) — static helper that wraps a
  plain params dict into the standard {'summary': DataFrame, 'estimates': dict} structure required by WalkForwardRunner.
  Non-MCMC adapters should call this in fit().
• BaseModel.fit() docstring updated to document the standardised
  return contract.

Bug fixes

• Fix _fit_window() treating all non-MCMC adapter fits as failures —
  'summary' key was absent from plain dict returns, causing every
  window to be skipped for GARCH, DL, HMM, Simple Breakout, MA, RSI.
• Fix GARCH z_score unit mismatch: (Close−Open)/(Open×GARCH_Vol)
  replaced with log_return×100 / GARCH_Vol — both now in the same
  log_return×100 scale.
• Fix GARCH NaN/inf errors: dropna() added before ARCH model fit.
• Fix HMM convergence failures: log_return × 100 scaling applied
  in both fit() and predict().
• Fix garch_btc.toml gamma_avg corrected from 2.70 → 2.0 to match
  revised z_score scale.
• Remove stale SNR fallback keys (alpha_long, alpha_short,
  sigma_1) from HMMRegimeAdapter.fit() and DlRegimeAdapter.fit()
  left over from removed build_dynamic_params().
• Fix _fit_window() empty-summary check: non-MCMC adapters using
  wrap_fit_result() produce an intentionally empty summary DataFrame.
  Previously this caused all non-MCMC windows to be skipped. Now only
  fails when both summary and estimates are empty.

Internal changes

• GarchCryptoAdapter now merges model-level [filters] config on
  top of shared backtest_settings.toml filters, allowing per-model
  filter overrides (e.g. use_sticky = false for GARCH).