Fix greedy portfolio allocation with integer prices

[DirkEmmett]

greedy_portfolio() crashes during weight normalization when quotes are stored as integer prices.

Use a floating-point calculation buffer without changing the caller's price Series or allocation policy. Adds regressions for integer and nullable-integer quotes, long/short portfolios, reinvestment, and budget conservation.

Validation: 317 tests passed, 4 skipped locally (Python 3.12, NumPy 1.26.4, pandas 2.2.3, SciPy 1.15.3). Ruff passed.

[arhancanli]

I hand-checked the new tests against the code on the PR branch, without running the suite.

• test_greedy_portfolio_integer_prices: first round buys A: int(50/30)=1 share (30) and B: int(50/20)=2 shares (40), leaving 30. Current weights are 30/70 and 40/70, so A has the larger deficit (0.5-0.4286 vs 0.5-0.5714). The second round buys A, leaving 0. That gives {A: 2, B: 2}, leftover 0, which matches the assertion.
• Short case, no reinvest: short_ratio follows the weights (0.25), so short_val=25 and long_val=100. Long A: int(100/30)=3, leftover 10. Short B: int(25/20)=1, leftover 5. Total leftover 15, allocation {A: 3, B: -1}, as asserted. With reinvest, long_val=125 gives 4 shares for 120 and leftover 5, plus 5 from the short side = 10, also matching.

The dtype=float on buy_prices is the line that avoids the in-place /= on an integer array, so it is the right place for the fix.

One thing the tests don't cover: lp_portfolio takes the same latest_prices. If it should accept integer or nullable-Int64 quotes too, a parametrised test there would show whether it already does or needs the same cast. I haven't run that path, so I don't know which.