hidden-Markov decoding of readout traces

New module decode: the inverse problem for the telegraph noise this package simulates - given a noisy readout trace, recover the occupation and the physical parameters.

• TelegraphHMM: two-state hidden Markov model in the exact rate convention of telegraph_traces (p_up = dt f/tauA, p_dn = dt (1-f)/tauA).
• • forward_backward / posterior: exact per-sample occupation posteriors (scaled forward-backward).
• • viterbi: most probable state path.
• • fit_hmm: Baum-Welch expectation-maximization estimate of (f, tauA, readout levels, noise) from the trace alone, with the EM monotone log-likelihood guarantee asserted in the tests.
    Validation against the package's own telegraph Monte Carlo: >99.9% state recovery at high SNR; the mean posterior confidence matches the realized accuracy to under a percent (the decoder knows how often it is right); EM recovers the ground-truth f and tauA within a few times the transition-count statistical error.

Also: README Status section reconciled with the release history.