Store combined transcript models in long form to bound peak memory

[dariansal]

Problem

combineSplicedTranscriptModels() builds one table carrying a start, end and readCount column per sample. A read class is observed in only a few samples, so that table is mostly NA, and the full_join fold materialises a second copy while the first is still live.

Peak memory therefore grows with rows × samples rather than with the data, and large sample sets can exhaust memory before the step completes.

Change

Keep only the populated (chain, sample) cells in long form and reduce them by chain.

The reduction also avoids expanding values by their weights. readCountWeightedMedian() repeats each value readCount times, takes the type-7 median and snaps up to the nearest observed value; with integer weights that is exactly the element at floor(N/2)+1 of the expanded sorted vector, locatable from the cumulative weights alone.

Output is unchanged

The existing implementation is kept as combineSplicedTranscriptModelsDense() and runs whenever the sparse path cannot guarantee an identical result:

• fewer than three samples, where nGroups is 1, no join happens, and the NSample* columns stay logical
• any input repeating an intron chain within one sample, which the join would fan out

The RNG draw stays where it was, so both paths see the same grouping for a given seed.

Verification

• upperMedianByGroup() against readCountWeightedMedian() across 50 randomised configurations
• both paths identical() on real read classes at n=20 and n=40, RNG seeded the same for each; the combine step ran 42.6× and 44.4× faster
• both paths against each other on synthetic read classes, plus the duplicate-chain fallback