Fix concurrent candle buffer reads

[saleh-mir]

Summary

• publish each DynamicNumpyArray buffer and logical index as one atomic state
• prepare appended rows and capacity before exposing them to concurrent readers
• use consistent snapshots for candle-service reads that need both the buffer and index
• apply the same behavior to single-row appends, batch appends, drops, flushes, and deletes

This prevents live readers from addressing a row that the underlying NumPy buffer has not allocated yet, and prevents drop operations from exposing mismatched array/index generations.

Verification

• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest tests/test_dynamic_numpy_array.py tests/test_candle_service.py -q — 18 passed
• deterministic concurrent growth-boundary harness for both append() and append_multiple()
• concurrency stress harnesses with 20,000 appends and four readers, including drop_at configurations
• python -m compileall -q jesse/libs/dynamic_numpy_array/__init__.py jesse/services/candle_service.py
• git diff --check

Notes

The local pytest environment has an incompatible auto-loaded anyio plugin (_pytest.scope is unavailable), so the focused suite was run with third-party plugin autoload disabled.