perf(data-layer): add bounded batch reads and Blob client reuse [2/5]

[Anthony Cintron (prbatero)]

Summary

Part 2 of 5 in the project layer-loading performance stack. Add reusable storage primitives that remove sequential amplification while enforcing one process-wide blocking-I/O budget.

Changes

• Add native Cosmos/PostgreSQL keyed batch reads and bounded fallback reads.
• Share one process-wide executor across metadata and artifact downloads.
• Reuse Blob service clients and lazily cached user-delegation keys.
• Match metadata types exactly so model cannot consume model_catalog.
• Make artifact downloads path-safe and atomic.
• Align read contracts across Blob, local filesystem, Cosmos DB, Data Lake, and PostgreSQL.

Testing

• Full hastelib suite: 553 passed.
• Black, isort, flake8, and detect-secrets passed.

Stack

• Depends on: perf(layer-loading): establish baseline and instrumentation [1/5] #185.
• This PR: 2/5, based on prbatero/feat/perf-layer-loading-baseline.
• Next branch: prbatero/feat/perf-project-details-api.

Review Notes

Review the diff against the configured base branch. Logical data-layer calls are not Azure REST transaction counts; Blob GET volume remains proportional to records returned.