Add LruFusionCache when direct bindings is used

[rdspring1]

This PR enables an LRU Fusion Cache decorator for direct bindings to address issue #2700.

Legacy Fusion Cache
iter 0: loss 0.1641, iter time: 102852.81ms, t: 4096
iter 1: loss 0.1196, iter time: 20309.14ms, t: 4096
iter 2: loss 0.0771, iter time: 19994.30ms, t: 4096

ToT Direct w/o LRU Cache
iter 0: loss 0.1641, iter time: 204575.74ms, t: 4096
iter 1: loss 0.1196, iter time: 22914.21ms, t: 4096
iter 2: loss 0.0771, iter time: 20006.35ms, t: 4096

Direct with LRU Cache
iter 0: loss 0.1641, iter time: 115281.74ms, t: 4096
iter 1: loss 0.1196, iter time: 20228.57ms, t: 4096
iter 2: loss 0.0771, iter time: 19998.50ms, t: 4096

• LRU Cache is 12% slower because equality check was updated to cover block
  scale operations, which create non-trivial allocation domains.

[kshitij12345]

LGTM, thanks @rdspring1

[kshitij12345]

Is this the minimum version which ships with LruFusionCache?

[rdspring1]

Yes

[kshitij12345]

Default value for max_fusions is 16384, we can remove this here.

[crcrpar]

qq: will there be an option of setting cache size? Or it would not be that useful

[rdspring1]

Usually we pick a reasonable number to avoid out-of-memory issues. We never thought to change it at runtime.