Concurrency safety fixes

[NHDaly]

• Fix thread-safety violation by accidentally not locking the base cache during delete!
• Fix concurrency-safety by not holding the get!() on the thread-local cache across the whole execution
• Use per-thread locks to make invariant to Task migration. ...... In retrospect, this does make this whole structure start to look a lot like a Dict + a Read/Write lock, and I wonder how their performance would compare......

Actually, this didn't seem to have too much perf impact, which is nice:

┌ Info: benchmark results
│   Threads.nthreads() = 1
│   time_serial = 0.03260747
│   time_parallel = 0.152217246
└   time_baseline = 0.131397639

┌ Info: benchmark results
│   Threads.nthreads() = 20
│   time_serial = 0.030174521
│   time_parallel = 0.307062643
└   time_baseline = 1.239406026

┌ Info: benchmark results
│   Threads.nthreads() = 100
│   time_serial = 0.030373533
│   time_parallel = 1.771401144
└   time_baseline = 5.397114559

[NHDaly]

Ugh the indentation change here down is annoying. I'd recommend hiding whitespace in the github UI: