10x faster task_local_memory

[chriselrod]

This is better than it was before, but I'll follow up with another PR to use stack-allocated memory in many cases. That'll probably have to wait until Monday.

As for this PR, the old version was slow because adding type asserts on a function means to call convert and then to type-assert, while adding it on a function return only means to type assert.

Getting from an IdDict{Any,Any} of course infers as Any, which then means we need a dynamic dispatch to the convert. By only having the type assert, we instead get a fast check to confirm the type and then an error.
I left the type assert on the function as well, but as the type has now already been confirmed by the time the function returns, we have a static dispatch to convert (that gets compiled away, because it's already known to be the correct type).

[codecov]

Codecov Report

Merging #94 (4b20efe) into main (3cda9f3) will not change coverage.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##             main      #94   +/-   ##
=======================================
  Coverage   71.41%   71.41%           
=======================================
  Files          13       13           
  Lines        2316     2316           
=======================================
  Hits         1654     1654           
  Misses        662      662           

Impacted Files	Coverage Δ
src/simple_chain.jl	87.09% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3cda9f3...4b20efe. Read the comment docs.