perf: generator performance optimization work

[guybedford]

This implements multiple and significant performance optimizations for the generator:

• Fetch pooling for browsers to avoid overloading the fetch API
• Increasing fetch retries from 3 to 5
• Ensuring fetch retries apply not only to the initial fetch call but also to the streaming of the response
• Coalescing all fetch calls consistently through the codebase to ensure there are never duplicate network requests in-flight. This is done carefully across both the trace map and provider implementations. In the process, the getTraceEntry method was refactored into the resolver implementation to share the cache with the exists call
• Optimizing the multi-pass visitation algorithm to remove the old looping convergence code - all installs are now simply two pass - trace + construct import map
• Optimizing the multi-install case when installing an array of install, so that phases are batched between all installs removing run-away looping behaviour for multi-installs

A test is added for a very large install case that didn't work at all when this work was started and now completes in 20 seconds.