[Operator Mechanism] Align cosine_similarity with torch2.12 (#79710)

* Align cosine_similarity with torch2.12

* fix

* Add accuracy compatible path to cosine_similarity

Keep torch's literal order, expand both inputs to the broadcast shape and
then take the norm, behind FLAGS_use_accuracy_compatible_kernel. Since
paddle.broadcast_to is not a 0-copy view like torch.expand, that path pays an
extra O(prod(broadcast_shape)) allocation per input, so the default stays on
unsqueeze plus the ||repeat(x, m)|| == sqrt(m) * ||x|| correction.

Tests cover the three shape families: equal shapes, mismatched ranks in both
dygraph and static graph, and large inputs broadcast along the reduced axis.
The peak memory test asserts the no-expansion property and is therefore
skipped when the compatible path is enabled.

* delete useless flag path in cosine_similarity

* Fix the broadcast factor under an unknown reduced axis

* Add detection for fp16/bf16 at cpu

* follow bots advice promote bf16/fp16 dtype to fp32 in cpu device

* Add test for cosine_similarity

* Fix the dtype promotion in cpu kernel.

Use fp32 as the accumulation_dtype when the common dtype is fp16/bf16 otherwise preserve the precision of the common dtype (in particular, do not demote a float64 input in a mixed reduced/float64 operation)

* Add test cases

* Fix cosine similarity related tests

* Fix p_norm_grad in processing x/0 to support test_zero_vector