Share body cache between harnesses within a codegen unit

[AlexanderPortland]

Improves on #4259 by fully sharing the body cache between harnesses in the same codegen unit.

This cache initially only stored bodies which had been modified by our transformations, with the goal being to avoid doing those transformations multiple times. It does this job well, having a hit rate of ~66% on queries for modified bodies¹.

However, I observed that even just querying to get the initial Body for a given Instance was a bottleneck, even if no passes needed to be applied. And, in this regard, the cache is much less effective, having a hit rate of only 1.8% if you count non-modified bodies¹.

Thus, this PR also adds all bodies to the cache, even those which were not modified by our passes.

Results

This change causes a further 7.7% reduction in the end to end compile time of the standard library (@ commit 177d0fd). The body cache has an average hit rate of 95.3% on that workload and, on a hit, we (at most) have to do an inexpensive clone from the cache rather than querying rustc_public.

As a caching solution, this comes with some memory overhead. Since we're sharing a cache and caching more bodies, the peak size of this cache has gone from 0.88 MB to 3.75 MB. This may seem like a significant increase (and I'm definitely open to feedback on whether it's acceptable), but given the fact that Kani hovered around 2.5-3.5 GB of RAM for that same run, this bigger cache would make up just 0.11%-0.15% of our total memory use.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 and MIT licenses.

Footnotes

1. when verifying the standard library (@ commit 177d0fd) ↩ ↩²