M7: Time series

Tracking issue for **M7**. Specification: everything tagged `(M7)` in [`06-pandas-parity.md`](../blob/main/docs/specs/06-pandas-parity.md).

### Scope

- [ ] `resample`, `asfreq` and upsampling
- [ ] `merge_asof` with backward, forward and nearest directions, `by` grouping and tolerance; plus `merge_ordered`
- [ ] Time-based rolling windows
- [ ] Full IANA timezone support with DST-aware arithmetic
- [ ] `date_range` and `bdate_range`
- [ ] The roughly forty pandas date offsets, plus business days and holiday calendars
- [ ] `at_time`, `between_time`, `infer_freq`
- [ ] Period and Interval types
- [ ] Sorted-key fast paths

### Exit criteria

- [ ] Every `(M7)` checkbox is ticked
- [ ] `merge_asof` matches pandas across a generated suite covering every combination of direction, tolerance and `by`
- [ ] The DST correctness suite passes across at least ten zones, with the explicit earliest / latest / raise policy asserted for ambiguous and nonexistent local times

### Why this milestone is worth its size

As-of joins and DST-correct resampling are simultaneously where pandas is most used and most error-prone. This is the milestone that earns adoption from finance and observability users specifically — the people for whom a wrong timestamp is not a cosmetic bug.

### Depends on

M6.

Read docs/specs/06-pandas-parity.md and resolve the M6 dependency first. Work through each item tagged (M7), using pandas as the comparison point. Done means every M7 checkbox is complete and the specified merge_asof generated suite and DST correctness suite pass.

Written by the indexing model from the issue text.