ci: cache ccache across build and reprotest runs

[j4n]

Every leg recompiles from scratch each run, and reprotest recompiles twice more on top of that. This commit introduces ccache for both jobs, keyed per distro/arch; reprotest reuses the build job's artifacts. Cache size capped at --max-size=500M for now, stats logged; with these on the first (cold) run:

Cacheable calls:   1800 / 2374 (75.82%)
  Hits:              13 / 1800 ( 0.72%)
  Misses:          1787 / 1800 (99.28%)
  Cache size (GB): 0.13 / 2.00 ( 6.47%)

[hpk42]

"caching is hard": this caching business is tricky and to be frank, i don't think it's worth optimizing/thinking much about it right now as long as compiling dovecot is a very rare task. It would need thorough testing to ensure a) it helps b) doesn't break anything on multiple runs. Doing cold-compiles is more predictable and reproducable so i'd just stay with that for now.

[j4n]

I was hoping it would shave off significant amounts of time from the new reprotest, but the results are too noisy across the few runs to support that hypothesis: At best, 44 vs 55 minutes, which is a long time either way; so closing this.

Job	No cache (07-30)	First cached run (08-10)	Rerun 1 (08-25 07:28)	Rerun 2 (08-25 09:00)	Δ (no cache → rerun 2)
bookworm/amd64 build	26m41s	25m19s	23m45s	24m49s	−1m52s
trixie/amd64 build	26m24s	25m40s	28m8s	24m11s	−2m13s
bookworm/arm64 build	19m28s	18m53s	19m39s	18m36s	−52s
trixie/arm64 build	19m12s	18m44s	19m57s	18m49s	−23s
bookworm/amd64 reprotest	44m12s	48m34s	45m0s	48m39s	+4m27s
trixie/amd64 reprotest	41m8s	33m18s	55m15s	44m45s	+3m37s

If we ever need to go back to this, we should probably export CCACHE_DISABLE=1 for the second reprotest run so its not using cached artifacts.