[Store] decouple DRAM decisions from metrics

[Hubert-Zhu]

Description

Implements a focused first slice of upstream RFC #3158.

• Add an authoritative DRAM usage snapshot derived from mounted allocators.
• Count the shared CXL allocator once without allocating on the 10 ms eviction path.
• Move memory eviction and promotion watermark decisions away from MasterMetricManager.
• Add allocator-state and metrics-reset regression coverage.

NoF eviction and the SsdMetricsProvider rename remain out of scope for this draft.

Module

• Mooncake Store (mooncake-store)

Type of Change

• Refactor

How Has This Been Tested?

Test commands:

git diff --cached --check
rg -n 'get_global_mem_used_ratio' mooncake-store/src mooncake-store/include
./scripts/code_format.sh --staged
/usr/bin/clang++ -std=c++20 -fsyntax-only -I./mooncake-store/include -I./mooncake-common/include mooncake-store/src/segment.cpp

Test results:

• Unit tests pass — not run locally because this macOS environment has no CMake or Boost.
• Integration tests pass (not applicable to this focused internal refactor)
• Static diff and residual-call-site checks pass.
• Formatting script was attempted but clang-format 20 is unavailable locally.
• Direct syntax compilation was attempted but stopped at the missing Boost headers.
• The fork currently exposes no enabled GitHub Actions workflows, so this Draft PR has no CI run yet.

Checklist

• I have performed a self-review of my own code
• I have formatted my code using ./scripts/code_format.sh (clang-format 20 unavailable)
• I have run pre-commit run --all-files and all hooks pass (pre-commit unavailable)
• Documentation is not required for this internal behavior-preserving refactor
• I have added tests to prove my changes are effective
• The existing RFC is #3158

AI Assistance Disclosure

• AI tools were used

Codex helped trace the allocator/metrics dependency, implement the focused change, and add tests. The human submitter should review and understand every changed line before any upstream submission.