feat: universal bank statement normalization layer (CSV, OFX/QFX, QIF)

[selenaalpha77-sketch]

Closes #112

Summary

Adds a universal bank statement normalization layer that converts diverse statement formats into FinMind's canonical expense schema.

New Endpoint

POST /expenses/import/normalize — accepts a file upload and returns normalised transactions with format metadata and deduplication fingerprints.

Supported Formats

Format	Extensions	Notes
CSV	.csv	Auto-detects Chase, HDFC, Axis bank profiles
OFX/QFX	.ofx, .qfx	Open Financial Exchange (SGML and XML variants)
QIF	.qif	Quicken Interchange Format

Bank Profile Auto-detection

The CSV parser inspects column headers to identify well-known bank layouts:

• Chase: Transaction Date, Post Date, Description, Category, Type, Amount
• HDFC: Date, Narration, Value Dat, Debit Amount, Credit Amount (split debit/credit columns)
• Axis: Tran Date, PARTICULARS, DR, CR, BAL (split debit/credit columns)
• Generic: best-guess fallback for any other CSV layout

Response Shape

{
  "format_detected": "CSV",
  "bank_profile": "Chase",
  "total_transactions": 42,
  "transactions": [{"date": "...", "amount": 25.00, "description": "...", ...}],
  "duplicate_fingerprints": ["sha256hex", ...]
}

Fingerprints are deterministic SHA-256 hashes of date|amount|description — callers can use them to detect duplicates before committing.

Files Changed

• New: packages/backend/app/services/bank_normalizer.py
• Updated: packages/backend/app/routes/expenses.py — added /import/normalize endpoint
• New: packages/backend/tests/test_bank_normalizer.py — 8 tests