v0.4.4

A patch. No published number changes and there is one new measurement, which is what firepanda's own planner costs against the same 43 queries written by hand.

pixi run clickbench-planner --size 1M runs all 43 twice in the firepanda driver over the same loaded table, once as the hand written port and once as the published SQL text through firepanda's front end, with parsing and planning inside the timed region. 33 of the 43 run both ways at 1M, the two routes agree on all 33, and the totals are 326.4 ms by hand against 543.9 ms planned, a factor of 1.67. TPC-H measured the same pair at 4.8, on a suite that has joins to order.

The pair was taken twice, hours apart on the same laptop, before and after the LIKE kernel landed in firepanda. Over the 29 queries both passes ran, each side fell by about 22 percent and the ratio moved from 1.615 to 1.617, so the ratio is what this measurement produces and the milliseconds are the machine's mood.

Full notes in CHANGELOG.md.