component: Defer host task creation

[adamrk]

Defer creation of a HostTask and insertion of it in the Store until it is required. This is an optimization on wasm->host calls where the HostTask may not actually be needed. The HostTask is materialized if the host future actually suspends (doesn't return Ready on the first poll), or if it tries to lower a borrowed resource (which requires inserting its CallContext into the Store to get a persistent identifier for it).

The work is separated into 3 commits:

• Add benchmarks for nop wasm->host calls that pass a borrow and where the host suspends before returning.
• Make the u32 identifier for CallContexts an proper enum
• The actual change of deferring HostTask creation.

These are the changes I see in the wasm->host benchmarks:
Baseline benchmarks:

component-async/wasm-to-host                                                                                          
                        time:   [354.02 ns 366.83 ns 380.05 ns]               
component-async/wasm-to-host-borrow              
                        time:   [462.76 ns 479.96 ns 497.66 ns]                                                       
component-async/wasm-to-host-pending-once                                                                             
                        time:   [2.5383 µs 2.6139 µs 2.6873 µs]

With changes:

component-async/wasm-to-host                                                                                          
                        time:   [283.83 ns 296.05 ns 308.65 ns]
                        change: [−21.162% −17.822% −14.374%] (p = 0.00 < 0.05)                                        
                        Performance has improved.        
component-async/wasm-to-host-borrow                                                                                   
                        time:   [377.80 ns 392.33 ns 406.01 ns]               
                        change: [−31.982% −29.587% −26.924%] (p = 0.00 < 0.05)
                        Performance has improved.          
component-async/wasm-to-host-pending-once                                                                             
                        time:   [2.4307 µs 2.5107 µs 2.5932 µs]                                                   
                        change: [−3.0505% +0.1437% +3.5901%] (p = 0.94 > 0.05)
                        No change in performance detected.

[alexcrichton]

Some great wins, thanks!

[alexcrichton]

Could you add a debug-assert on this branch that materialized_task is None?

[alexcrichton]

Could this be moved into host_task_create? To the extent that set_thread is a somewhat-hot function and I think that's the only caller which would otherwise trigger this, it might help keep this function a bit leaner to move it over there. That might mess up and/or require moving around some debug assertions though

[alexcrichton]

And, similarly, for handling deferred_host_call_context getting cleared above, could that move into the first_poll function or similar?