Make hat allocations more stable

[pokey]

The problem

Today, we don't make any effort to use the same hats for the same tokens when we refresh the hat map. This behaviour can lead to problems where

• the hat map can shift while you're thinking / formulating a command,
• it can be distracting
• you cannot start issuing a new cursorless command while the previous one is running
• race conditions with scrolling / edits while you're starting to say a command

The issue was improved by #318, but that only works for single phrases, not continuity between phrases

The solution

We could treat hats more like a resource, the way a memory allocator treats memory blocks. Then when we go to reallocate the hat map, we only allocate new hats, or hats that have been freed from a deletion / scroll

Note: a lot of implementation might overlap with #308, and arguably this issue is a better solution to the problem that #308 attempts to solve, though #308 would allow frozen sections to persist even when off screen

Fwiw, here is some initial work: #1225 / a4802ea