fix(data): make FileInstrumentStorage round-trip instrument files correctly

[vignesh-nagarajan-vn]

Two long-standing defects in FileInstrumentStorage broke the instrument file round-trip:

• _read_instrument let pandas NA-parse the symbol column, so real tickers literally named "NA" or "NULL" were loaded as float NaN and merged into a single bogus instrument, which then leaked out of D.list_instruments. PR Fix the bug of reading string NA as NaN in the function exists_qlib_data. #1736 fixed this same defect in exists_qlib_data (reported in Ticker name "NA" makes the exists_qlib_data function report errors.  #1720), but the storage reader - the path that actually feeds data loading - kept the old behavior.

• _write_instrument wrote the file twice back-to-back; the second to_csv immediately overwrote the correctly-ordered output with columns in [start, end, symbol] order, which _read_instrument then misparsed (dates became symbols). Present since the original implementation (70fc581); every public mutation (update, __setitem__, __delitem__) went through it.

Fix: disable default NA parsing when reading the symbol column (symbols are opaque identifiers), and write the file once in the documented [symbol, start, end] order, keeping the utf-8 encoding of the removed call.

Both defects are covered by new self-contained tests in tests/storage_tests/test_storage.py (temporary provider, no market data), each verified to fail against the previous code.

Behavior for well-formed data is unchanged: the existing tests/storage_tests pass as-is, D.list_instruments output on qlib_data_simple is hash-identical before/after the change, and black/flake8/pylint/mypy all pass with the CI invocations.

Fixes #1720 (completes the fix that #1736 applied only to exists_qlib_data)

[vignesh-nagarajan-vn]

@microsoft-github-policy-service agree