Increase gRPC max recv message size in Bazel client

[clintharrison]

We have seen invocations where the FailureDetail [presumably] is larger than the 4MiB default (we've seen this when we have many CacheNotFoundException exceptions; the list of missing digests is very very large, and the associated stack traces are even larger).

This causes the Bazel client to exit with a misleading error:

Server terminated abruptly (error code: 8, error message: 'Received message larger than max (9620141 vs. 4194304)', log file: '/mnt/tmpfs/BAZEL_OUTPUT/server/jvm.out')

Somewhat confusingly, the server has not actually terminated, but it's a client-side enforced RESOURCE_EXHAUSTED.

I picked 20 MiB pretty arbitrarily, it's just what happened to be large enough for us.

[fmeum]

@coeuvre for the long digest lists

[tjgq]

I don't see an issue with increasing the maximum response size, but I think we should (also) improve the error message and stack trace for CacheNotFoundException; reporting every single missing digest, and a separate stack trace per digest, is unlikely to be useful.

[ulrfa]

The pull request #23490 reduces the length of BulkTransferException error messages and avoids mentioning all missing digests. @tjgq or @coeuvre, please let me know if you would like me to also try removing stack traces.

[clintharrison]

Thanks, that's a much nicer fix I didn't know how to make 😆