Passing enable_gqa flag in attention dispatcher should be version guarded

[a-r-r-o-w]

Just realized this is potentially breaking:

diffusers/src/diffusers/models/attention_dispatch.py

Line 229 in cde02b0

"enable_gqa": enable_gqa,

I believe enable_gqa was first released in PT 2.5.0. For versions before that, this flag should not be passed. Additionally, some user code may have wrappers around SDPA and do not expect the enable_gqa flag (think of TorchFunctionMode being used to intercept call).