Set amax_and_scale_synced unconditionally

[awgu]

Stack from ghstack (oldest at bottom):

• Removed module arg from fsdp_pre_all_gather #217
• [POC] Added fp8 all-gather extensions #216
• Added initial compile tests for flat-parameter FSDP #215
• Added use_activation_hooks: bool to swap #214
• Used functional all-reduce for amax reduction #219
• -> Set amax_and_scale_synced unconditionally #220

Float8Linear.amax_and_scale_synced is orthogonal to config.enable_amax_init.

• enable_amax_init = False skips the lazy initialization of amaxes and scales. Without this, the 1st iteration's amaxes will be the default (E4M3_MAX_POS).
• amax_and_scale_synced is checked every pre-forward regardless of config.enable_amax_init.

float8_experimental/float8_experimental/float8_linear.py

Lines 266 to 273 in 956195b

	if (
	self.is_amax_initialized
	and (not self.amax_and_scale_synced)
	and torch.is_grad_enabled()
	):
	raise AssertionError(
	"amaxes and scales not synced, please call `sync_float8_amax_and_scale_history` before forward"
	)

[drisspg]

So I added this because this module mutation was causing a graph break, hmmm

[awgu]

@drisspg Would it be possible to define an inner function and compile that? That might also help with any other graph break issues, e.g. like the fp8_layers comment.