lib-cuda: record the measured cost of the 32-bit index axis on this backend

[balbasty]

Comment-only follow-up to #143, which merged before these numbers were in hand.
No build behaviour changes.

#143 landed INDEXFLAGS / FF_INDEX32 and made build-cuda a two-leg matrix
over the two positions. Its own text still quoted the axis's cost from the
host-compiler measurement in #94 plus an assumption that CUDA would behave
similarly. The matrix's first run has now produced the real CUDA figures, so
this records them above MODULES in src/lib-cuda/Makefile — both positions,
one CI run, one runner class, nvcc -O1 -j2 with the shipping bound/spline
policy.

Peak nvcc RSS (GiB) / wall per module:

module	FF_INDEX32=1	FF_INDEX32=0	Δ RSS
reg_flow	12.97 / 1023 s	5.69 / 497 s	−56.1%
reg_field	8.09 / 791 s	3.76 / 343 s	−53.5%
reg_field_rls	7.07 / 701 s	3.42 / 353 s	−51.7%
resize	2.00 / 193 s	1.08 / 113 s	−46.2%
reg_flow_rls	1.90 / 155 s	1.01 / 89 s	−46.9%
pushpull_backward	1.51 / 167 s	0.80 / 102 s	−46.9%
pushpull	1.48 / 158 s	0.79 / 97 s	−46.6%
restrict	1.30 / 120 s	0.68 / 72 s	−47.4%
distance	0.80 / 118 s	0.54 / 61 s	−33.1%
splinc	0.42 / 68 s	0.27 / 37 s	−35.5%
posdef	0.37 / 40 s	0.35 / 28 s	−6.9%
total nvcc CPU	3532 s	1791 s	−49.3%
job wall clock	~38 min	~19 min

Source: run
32363152580,
jobs build-cuda (index32) and build-cuda (index64), both green.

Two things this makes visible, and the reason for recording rather than
quietly updating:

• CUDA is worse than the host-compiler extrapolation implied, not better.
  reg_flow sheds 56.1% of peak RSS here, against the 44.6% clang showed.
• The existing table's reproducibility claim does not hold uniformly. It
  says the heavy modules reproduce to better than 0.1%. reg_flow does — 12.97
  against the recorded 12.98 GiB. reg_field came in at 8.09 against a
  recorded 8.93, ~9% low. Both figures are now in the file with that noted,
  instead of the older one being overwritten as though nothing had moved.

The -j2 overlap hazard described further down that comment is unchanged:
halving each peak halves the sums too, but only in the position that is not the
default.

MIGRATION.md gets one correction in the same commit: it said "365 of 365
relevant impl/cuda upload sites use the synchronous copyToDevice", a figure
inherited from #94. On current main it is 363 of 391 — distance_ {euclidean,l1,mesh}.h now use copyToDeviceAsync. The point stands for almost
every op; the number was stale.

The benchmark that would actually settle the axis is still #144.

---

Generated by Claude Code