Fix resume failing when interrupt checkpoint is buried by continuation checkpoints

[MikaAK]

Summary

• When resuming from an interrupt, the continuation may save intermediate checkpoints that become the "latest" by created_at
• A subsequent resume call loads only the latest checkpoint via load/1, finds no pending_interrupts, and returns {:error, :no_pending_interrupt}
• This fix adds a fallback in resume_from_checkpoint/6 that uses list/2 to search recent checkpoints for the most recent one with pending interrupts

Scenario

1. Graph invokes → hits interrupt at node A → checkpoint saved with pending_interrupts
2. Resume resolves A → graph continues through B → C
3. Node B saves a checkpoint (now the latest, no interrupts)
4. Node C hits another interrupt → checkpoint saved with pending_interrupts
5. Before fix: Second resume loads B's checkpoint (latest), sees no interrupts → fails
6. After fix: Second resume falls back to searching checkpoint list → finds C's interrupt → resumes correctly

Test plan

• Added test resume finds interrupt checkpoint even when later non-interrupt checkpoints exist covering the multi-interrupt resume scenario
• All 144 tests pass (143 existing + 1 new)
• No changes to the Checkpointer behaviour — uses existing list/2 callback

[twist900]

@claude review