Reduce size of in-memory benchmark caches

[ChrisCummins]

• Limit the size of the LLVM benchmark cache at 128 items, with random eviction.
• Reduce the size of the in-memory Benchmark protocol buffer cache from 512MB to 256MB.
• Change the type of the StartSessionRequest.benchmark field from a string to a Benchmark message.
• Add an always_send_benchmark_on_reset option to ConnectionSettings that forces the environment to always send the full Benchmark data to the backend service on every reset(). This is useful for cases where benchmarks are rarely repeated, e.g. for training loops with very large datasets.
• Only run LLVM initialization once.
• Some small improvements to service logging.

[codecov-commenter]

Codecov Report

Merging #299 (4fafbe6) into development (89c7897) will decrease coverage by 0.02%.
The diff coverage is 61.53%.

@@               Coverage Diff               @@
##           development     #299      +/-   ##
===============================================
- Coverage        82.43%   82.41%   -0.03%     
===============================================
  Files               87       87              
  Lines             4766     4776      +10     
===============================================
+ Hits              3929     3936       +7     
- Misses             837      840       +3     

Impacted Files	Coverage Δ
...mpiler_gym/service/runtime/compiler_gym_service.py	21.56% <0.00%> (-0.44%)	⬇️
compiler_gym/envs/compiler_env.py	85.78% <85.71%> (-0.01%)	⬇️
compiler_gym/service/connection.py	74.11% <100.00%> (+0.09%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 89c7897...4fafbe6. Read the comment docs.