fix(retry): avoid truncating fractional retry_percent in TpsBudget

[SAY-5]

TpsBudget::new used (1, (1.0 / retry_percent) as isize) for retry_percent <= 1.0, which truncates the reciprocal so e.g. 0.6 authorizes a full retry per deposit instead of 0.6. This reuses the existing 1000x deposit scaling for all positive percentages and adds a regression test. Closes #857.

[seanmonstar]

I think most often the percent will be 10, 20, around there. So, to not need to deal with large numbers in the most common case, this could probably be edit to just <= 0.5. What do you think?

[SAY-5]

Makes sense, done. Anything at or below 0.5 keeps the old (1, 1/percent) pair, and only above that does it scale by 1000, since that's where the reciprocal truncates to 1. Worth noting the small values still truncate (0.3 gives 3, so effectively 33%), but that's the existing behavior and the numbers stay small.

[SAY-5]

Good call, switched to scaling only when retry_percent <= 0.5 so the common integer-ish percents keep the original arithmetic. Updated the test name to match.

[SAY-5]

Updated: applied the 1000x scaling uniformly for all non-zero retry_percent values (not just <= 0.5). This means the test case at 0.6 now actually verifies the fix: (1000, 1666) gives exactly 6 retries per 10 deposits. The > 1 case also simplifies to the same formula.