Preliminary STM benchmarks and low-hanging fruit optimizations

[jdegoes]

A few STM benchmarks and a variety of performance improvements, including:

• Addressing performance regression caused by ZIO environment (the cost of provide was dominating short-lived benchmarks due to the use of finalizers; now the top-level environment is "free")
• Refactored STM to use java.util.HashMap, mainly because it is easier to minimize allocations compared to Scala's mutable map (iteration through a Scala map will always allocate tuples)
• Introduced specializations for effect* variants
• Make ensuringR cost-free, which means it became ensuring, which allowed elimination of environmental access in bracket
• As a result of other cost-free changes, made more operators polymorphic in R
• bracketExit is now fast enough (no forking) it can be the sole implementation

There's much more work to be done, but the runtime system is generating very clean, allocation-free code; STM has been through "first-pass" optimization; and overall, things are looking cleaner and more uniform thanks to the cost-free refactorings.

The main learning of STM benchmarks is that TSemaphore is as fast as the hand-written ZIO Semaphore; and TQueue is as fast as FS2 Queue in parallel access, although only 50% as fast in two other scenarios. With further optimization, it seems likely STM can produce performance competitive with hand-written, hand-optimized Ref + Promise code.

[wi101]

Looks great!