Parse samplers on demand

[HansKristian-Work]

Fixes #196.

Before on the reported archive, there were millions of sampler objects. Memory usage exploded:

Fossilize INFO: =================
Fossilize INFO:  Memory usage:
Fossilize INFO:    #0:   367 MiB resident     3 MiB shared (0 MiB shared metadata).
Fossilize INFO:    #1:  2186 MiB resident   237 MiB shared (242 MiB shared metadata).
Fossilize INFO:    #2:  2156 MiB resident   238 MiB shared (242 MiB shared metadata).
Fossilize INFO:    #3:  2168 MiB resident   238 MiB shared (242 MiB shared metadata).
Fossilize INFO:    #4:  2013 MiB resident   238 MiB shared (242 MiB shared metadata).
Fossilize INFO:  Dirty filesystem writes: 0 MiB.
Fossilize INFO:  IO stall: 0%.
Fossilize INFO:  Num running child processes: 4.
Fossilize INFO: =================

On a pruned archive, there is only 1 surviving sampler, so we get:

Fossilize INFO: =================
Fossilize INFO:  Memory usage:
Fossilize INFO:    #0:   137 MiB resident     3 MiB shared (0 MiB shared metadata).
Fossilize INFO:    #1:   284 MiB resident    89 MiB shared (88 MiB shared metadata).
Fossilize INFO:    #2:   289 MiB resident    89 MiB shared (88 MiB shared metadata).
Fossilize INFO:    #3:   285 MiB resident    89 MiB shared (88 MiB shared metadata).
Fossilize INFO:    #4:   286 MiB resident    89 MiB shared (88 MiB shared metadata).
Fossilize INFO:  Dirty filesystem writes: 9 MiB.
Fossilize INFO:  IO stall: 0%.
Fossilize INFO:  Num running child processes: 4.
Fossilize INFO: =================

This is closer to what we normally expect as a high water mark per process.

With this PR, VkSamplers are only replayed if they are a dependency of a descriptor set layout. Unfortunately, we have a hard time knowing which samplers to record since we don't know at recording time if a sampler is part of a descriptor set layout. We don't want to bloat the runtime memory usage by holding onto an unbound number of samplers "just in case". I'll follow up with another PR to be able to ignore immutable samplers if we know the application (DXVK in this case) does not really make use of them.

Fossilize INFO: =================
Fossilize INFO:  Memory usage:
Fossilize INFO:    #0:   368 MiB resident     3 MiB shared (0 MiB shared metadata).
Fossilize INFO:    #1:   516 MiB resident    90 MiB shared (242 MiB shared metadata).
Fossilize INFO:    #2:   520 MiB resident    89 MiB shared (242 MiB shared metadata).
Fossilize INFO:    #3:   516 MiB resident    89 MiB shared (242 MiB shared metadata).
Fossilize INFO:    #4:   518 MiB resident    89 MiB shared (242 MiB shared metadata).
Fossilize INFO:  Dirty filesystem writes: 18 MiB.
Fossilize INFO:  IO stall: 0%.
Fossilize INFO:  Num running child processes: 4.
Fossilize INFO: =================

This is far more reasonable, but the increased memory usage compared to pruned is the hashmap post-parse needs to holds millions of sampler references. Not recording millions of samplers will fix this as well down the line.

[Frogging101]

Great!