Forgotten cards should stay due the same day, not roll to tomorrow

[oddish3]

MIN_INTERVAL (1 day) is applied to every grade, including Forgot. In practice this means grading a card "forgot" always pushes its due_date to tomorrow — even for a brand-new card, whose raw FSRS interval on Forgot is typically well under a day (~0.4 days).

hashcards has no Anki-style intraday relearning steps (by design, per the "Simple" principle in the README), so the only way a forgotten card can come back up in a later session on the same day is if its due_date is allowed to stay today. Right now it can't — you have to wait until tomorrow, which doesn't match what "Forgot" should mean.

This adds a separate, lower floor (MIN_INTERVAL_FORGOT = 0.0) used only for Grade::Forgot, leaving Hard/Good/Easy untouched. It's a one-constant, one-branch change — no session/queue changes, no schema changes, still fully day-granular (no new intraday state). It composes naturally with all_due(today) (used by drill and, after #199, by due), so a forgotten card just shows up as due again whenever you next open a session that day.

Added three tests: forgetting a new card, forgetting a previously-reviewed card, and forgetting the same card twice same-day (checking elapsed_days == 0 doesn't break the stability update - FSRS handles that as an ordinary case, not a special one).

Happy to adjust the floor value if you'd rather it not be exactly 0 (e.g. leaving a small epsilon), just picked 0 since nothing downstream treats interval_days as a denominator.

[oddish3]

Problem

Grade::Forgot was subject to the one-day minimum interval. If a user ended a
session after forgetting a card, that card could not appear in another session
until tomorrow.

Change

For Grade::Forgot, persist interval_days = 0 and due_date = today, while
retaining the true FSRS result in interval_raw. Other grades are unchanged.

Result

• Same-session repetition is unchanged and remains queue-controlled.
• A forgotten card is eligible in another session that day.
• If not reviewed again, it remains eligible as overdue on later days.
• There are no intraday delays or relearning steps.
• Future due reports retain exact-date semantics.

Added a database-level regression test covering persistence and all_due.
Verified the real HTTP lifecycle across three sessions. All 149 tests pass.