Close incorrect quadratic-memory fix for u_sf_mRNA_factors

[TTTPOB]

Correction

This PR is being closed because its memory justification was incorrect. The (B, 1, B, F) Delta.v is a zero-stride expanded view that shares the original (B, F) mRNA storage; its logical numel() does not represent physical allocation.

pyro.deterministic() uses .mask(False), and Pyro's MaskedDistribution.log_prob() returns an expanded scalar zero without evaluating or materializing the underlying Delta value. The inspected TraceELBO path only consumes the resulting scalar log_prob_sum.

Moreover, event_dim=0 gives the Delta a (B, F) batch shape, which causes Delta.__init__ to allocate a real (B, F) log_density tensor. The proposed change therefore does not eliminate quadratic physical storage and can add a linear allocation.

The regression tests correctly identify a change in logical distribution shape, but they do not demonstrate reduced VRAM. No Cell2location code change is justified by the evidence presented here.

Detailed correction and measurements: #444

[TTTPOB]

Closing after correcting the allocation analysis in #444. The expanded Delta is a zero-stride view rather than quadratic physical storage, and event_dim=0 can add a linear log_density allocation.