linker: batch mem2reg to process all variables in a single pass

mem2reg processed each variable independently.
This O(V*N) complexity caused multi-minute hangs on large post-inline
functions.

Batch all operations into single passes over the instructions.

Fixes #546. Measured on the proof-of-space-gpu shader:
mem2reg drops from 228s to 18s (13x), total link from 236s to 26s.

Also adds unit tests.

Author: LegNeato