Add LongCat-Flash-Lite-Sparse (LSA + n-gram) (#2063)

* Add LongCat-Flash-Lite-Sparse (LSA) model support

* Add LongCat-Flash-Lite-Sparse MTP speculative drafter

* Add LongCat-Flash-Lite-Sparse target-side speculative verify hooks

* Optimize LongCat-Flash-Lite-Sparse: fuse n-gram projection, reuse indexer epilogue kernel

* Fix LongCat-Flash-Lite-Sparse n-gram fusion: keep word embedding at full scale

LongcatCausalLM fuses word + sum(ngram_projs)/(1+num_embedders), keeping the
word embedding at full scale, unlike LongcatFlashNgram which divides the whole
sum. Dividing the word by 13 garbled real-weight generation; this restores
coherent output (verified on the 4-bit Lite-Sparse checkpoint).

* Remove MTP speculative decoding from LongCat-Flash-Lite-Sparse

Batch-1 self-speculation did not beat launch-bound decode for this model
(~parity at short/mid context, worse at long context), so drop the MTP
drafter, exact verifier, and the model's speculative-decoding hooks. The
core LSA + n-gram model support is unchanged.

* Add LongCat-Flash-Lite-Sparse model README

* Update README.md

Signed-off-by: Alazer Manakelew <alazermanakelewb@gmail.com>

* edit links

Signed-off-by: Alazer Manakelew <alazermanakelewb@gmail.com>

* Fix usage load path to mlx-community in LongCat README

---------

Signed-off-by: Alazer Manakelew <alazermanakelewb@gmail.com>