This commit solves a number of bugs and adds some caching related
functionality. Specifically:
1. Caching is now input-based. In cases of completely identical nodes,
the output will be reused (for example, if you have multiple
LoadCheckpoint nodes loading the same checkpoint). If a node doesn't
want this behavior (e.g. a `RandomInteger` node, it should set
`NOT_IDEMPOTENT = True`.
2. This means that nodes within a component will now be cached and will
only change if the input actually changes. Note that types that can't
be hashed by default will always count as changed (though the
component itself will only expand if one of its inputs changes).
3. A new LRU caching strategy is now available by starting with
`--cache-lru 100`. With this strategy, in addition to the latest
workflow being cached, up to N (100 in the example) node outputs will
be retained. This allows users to work on multiple workflows or
experiment with different inputs without losing the benefits of
caching (at the cost of more RAM and VRAM). I intentionally left some
additional debug print statements in for this strategy for the
moment.
4. A new endpoint `/debugcache` has been temporarily added to assist
with tracking down issues people encounter. It allows you to browse
the contents of the cache.
5. Outputs from ephemeral nodes will now be communicated to the
front-end with both the ephemeral node id, the 'parent' node id, and
the 'display' node id. The front-end has been updated to deal with
this.
Control loras are controlnets where some of the weights are stored in
"lora" format: an up and a down low rank matrice that when multiplied
together and added to the unet weight give the controlnet weight.
This allows a much smaller memory footprint depending on the rank of the
matrices.
These controlnets are used just like regular ones.
