Bucket cleanup

[SirTyson]

Description

Resolves #4031

This PR includes cleanups and hardening of Bucket related code.

• loadPoolShareTrustlinesByAccountAndAsset: Extended a test to make sure this function does not return outdated entries. This was working properly, but did not have an explicit test. I've also added an explicit seenKeys set instead of relying on the fact that unordered_map::emplace will not update a key if it already exists for more defensive programming. Finally, there was an edge case where if a Bucket only contained offers and did not contain a METAENTRY, the scan would return early. This should never happen in practice, but it has been fixed.
• XDRStream audit: I took a deep dive into ifstream and ran several experiments such as disabling stream buffering. I've concluded that ifstream with buffering is still the optimal file IO interface for BucketsDB. I've also checked that we're handling EOF and bad bits properly, and could not find any dependencies on out-of-spec behavior. The only code change was a minor scoping issue on a tracy metric.
• Added separate stream for BucketDB scans and eviction scans.
• Warn more aggressively if eviction scans get stuck. Previously, there was only a single warning per stuck bucket, and there we certain edge cases where the warning would not be triggered, such as if a large bucket was almost entirely TEMPORARY entries. Now, we warn on every ledger and check much more aggressively.

Checklist

• Reviewed the contributing document
• Rebased on top of master (no merge commits)
• Ran clang-format v8.0.0 (via make format or the Visual Studio extension)
• Compiles
• Ran all tests
• If change impacts performance, include supporting evidence per the performance document

[marta-lokhova]

r+ 4f0789d