Add return_raw option to BlackLittermanModel.bl_weights

[heykav]

Closes #763.

What

bl_weights() normalizes the solved weights to sum to 1 before returning them. That's the right default, but it discards information: the raw (pre-normalization) solution's total encodes how much net borrowing/lending the posterior implies, and there's no way to recover it once it's been normalized away.

Adds return_raw=False to bl_weights(). When True, it returns the raw weights through the same _make_output_weights() path instead of the normalized ones.

Why this is safe

self.weights is always set to the normalized version regardless of the flag — only the return value changes. optimize() (the alias) and every other caller that reads self.weights afterward (e.g. portfolio_performance()) sees identical behavior to before this change.

Verification

• New regression test test_bl_weights_return_raw: confirms the raw weights are proportional to the normalized weights by exactly the sum that normalization divided out, confirms self.weights stays normalized, and confirms the default (return_raw=False) call is byte-identical to pre-change behavior.
• Full suite: 280 passed, 32 skipped. The one failure in test_hrp.py::test_hrp_errors is the pre-existing, already-tracked scipy._LINKAGE_METHODS incompatibility (HRPOpt fails with current SciPy because _LINKAGE_METHODS was removed #754, already has an open PR at Fix HRP SciPy linkage method validation #755) — unrelated to this change, reproduces identically on main without it.
• ruff check and black --check clean on both changed files.

[heykav]

Noting for the reviewer: @jrvarma suggested on #763 that this might read better as normalize=True (returning raw weights when False) rather than return_raw=False. Both are reasonable — there's no existing normalize= convention elsewhere in this codebase to anchor to, so I went with the more literal return_raw naming, but happy to switch if that's the preferred convention here. Let me know either way and I'll update it in this PR.