Detailed size estimation + fix #193 (overestimating fused QKV storage)

[mratsim]

This PR does 2 things related to size:

1. Fix [Bug] storage_size (TP) overestimates allocation of fused linear layers like QKV / QKVZ #193, for fused QKV layers, when key_q, key_k, key_v are not set, self.key is undefined and storage will used the FP16 formula here:

   exllamav3/exllamav3/modules/linear.py

   Lines 430 to 463 in cb1a436

   	@cached_property
   	def _storage_size(self):
   	# alt_key is only used when loading unquantized model
   	if self.is_exl3_storage(self.key):
   	return sum(self.config.stc.get_tensor_sizes(prefix = self.key))
   	else:
   	return 2 * self.in_features * self.out_features
   	def storage_size(self):
   	return self._storage_size
   	def recons_size(self):
   	return 2 * self.in_features * self.out_features
   	def make_tp_allocation(self, options: dict) -> list[TPAllocation]:
   	storage = 0
   	storage += self.storage_size()
   	overhead_d = self.out_features * (self.out_dtype or torch.half).itemsize
   	overhead_s = self.out_features * (self.out_dtype or torch.half).itemsize
   	recons = self.recons_size()
   	tpa = TPAllocation(
   	key = self.key,
   	channel_width = 128,
   	channel_unit = "channels",
   	storage_per_device = 0,
   	storage_to_split = storage,
   	overhead_per_device = overhead_d,
   	overhead_to_split = overhead_s,
   	recons_temp = recons,
   	channels_to_split = self.out_features // 128,
   	limit_key = "linear"
   	)
   	return [tpa]

   This can be fixed simply by setting those keys at the model architecture level.

2. This expands model.py with get_detailed_weights_info that is then used by size_estimation.py for detailed statistics.

I noticed 2 things:

• Convolution has no weights_numel set if weights are not loaded so I added a formula to compute it from metadata.
• Linear doesn't count the number of elements in the bias, that might affect TP allocation and might be a bug. I don't count the number of elements in convolution bias to mirror Linear layers.

[mratsim]

The output of size_estimation.py looks like this now