Repair the unit-test suite and wire it into CI

[nh13]

The crate's #[cfg(test)] modules had stopped compiling — they still used the old ModelData API and by-value ModelInput conversion, so cargo test failed with ~50 errors and none of the unit tests ran. Since CI only runs the tests/ integration Makefile (cargo build, not cargo test), this went unnoticed. This migrates every test to RMITrainingData / &ModelInput, rewrites the two mod.rs tests that used removed helpers against set_scale/iter, and adds a cargo test -p rmi_lib step to .drone.yml so they keep running. (Bare cargo test at the repo root only exercises the rmi binary, since rmi_lib is a path dependency.)

Reviving the tests surfaced one real bug: FixDupsIter (behind RMITrainingData::iter()) re-emitted the final training row, so every model saw one phantom duplicate. The CDF trainers (ncdf/lncdf/loglinear_slr) compute statistics with n = len() while iterating len()+1 times, which skewed their parameters; stopping the iterator instead of replaying corrects them. Heads-up: this slightly changes generated-RMI training output, so it's worth a look under the integration suite (I couldn't run the 200M-key SOSD benchmarks locally) — test_iter now pins the iterator's behavior.

One expectation was stale rather than buggy: the equidepth-histogram test expected 333 for a key past all training data, a leftover from the old PLR-based linear_norm_histogram. The current equidepth_histogram has num_bins = 333, so the max bin index is 332; I corrected the expected value.

Made in collaboration with Claude.