Creating a query heap per query might be too expensive

[jenatali]

Trying to run Tracy through D3D11On12 takes far too long to start up, with pretty much all of the time spent creating 64*1024 timestamp queries. Each of these timestamp queries creates a 4-query heap, which likely involves allocating a 64KiB query heap to store 32B.

This should probably turn into a device (immediate context) level query heap pool per type where query heap slots can be suballocated out to individual queries.