Add forward and backward transfer metrics

[anagainaru]

Adds eval/fwt and eval/bwt to the metrics logged for a run, so a run can answer whether continual learning is actually helping the window it fired on, and whether it's paying for that by forgetting earlier ones.

Without this PR a run reports eval/test_curr_acc and eval/test_hist_acc. test_hist_acc is a single pooled score over concatenated history with no reference point, so it conflates forgetting with windows that were simply harder. The transfer metrics fix that by scoring each past task against its own starting point.

What's logged

A task is one drift event — the window the detector fired on and the CL loop adapted to — so T counts adaptations, not stream windows. Both metrics are entries of the train-test matrix R, where R[i][j] is the score on task j after finishing task i:

Metric	Definition	Reads
eval/fwt	R[i][i] - R[i-1][i]	The gain adapting delivered on the triggering window. Logged at every drift event, first included.
eval/bwt	(1/(T-1)) · Σ_{i<T} (R[T][i] - R[i][i])	Each past task scored now, minus its score right after it was learned. Absent on the first drift event, where the sum is empty.
eval/test_pre_cl_acc	R[i-1][i]	The pre-CL score, logged directly so fwt is reconstructible from the CSV.

Sign convention

Both metrics are raw differences of R, so the sign inherits the metric's direction. With accuracy (higher better) negative bwt means forgetting; with a lower-is-better metric such as SLAC-FEL's MAE, positive bwt means forgetting. Check the harness's higher_is_better before comparing across examples.

One trap worth noting: positive bwt on accuracy can mean real backward transfer or a weak R[i][i] to begin with — a short train.max_iter leaves the diagonal undertrained and manufactures meaningless positive values. Sanity-check eval/test_curr_acc per event first.

This fwt is the CL gain on the triggering task, not the Lopez-Paz & Ranzato form R[i-1][i] - b_i; no baseline model term is computed.

Example output

For MNIST:

[krm9c]

Okay, the metric calculation looks correct and the mnist plots look correct.

[krm9c]

I think, the plots and the calculation looks correct to me.