Yishuo Wang

## Description When setting `LD_PRELOAD`, the `bigdl-nano-init` script prefers to look for `libtcmalloc.so` and `libjemalloc.so` in `$LIB_DIR` directory, which is `${CONDA_PREFIX}/lib`. However, these .so files should be in `${NANO_DIR}/libs` directory....

## Description Add zsh support for `bigdl-nano-init` script fix #5751

openvino and intel-tensorflow>2.7 are incompatible

11

When importing both openvino and intel-tensorflow > 2.7, the following code will raise an error: `RuntimeError: dimension (-1) in node dim must be greater or equal to -1: at offset`...

Calling `bigdl.nano.tf.keras.Model`'s `quantize` method may cause error when using tcmalloc: ``` tcmalloc/tcmalloc.cc:1726] size check failed 4096 8 1 tcmalloc/tcmalloc.cc:2103] CorrectSize(p, size, DefaultAlignPolicy()) @ 0x7f90d0dbfc0d 0x7f90c2928634 0x7f90b99a3e98 0x7f90b998a561 0x7f90bef02f00 0x7f90be29fdf7 0x7f90bef02052...

set jemalloc as the default allocator

1

## Description * set jemalloc as the default allocator

A better `packNibbles` and `mul_sum_i8_pairs_float` implementation using AVX512

2

Use only three instructions to implement `packNibbles` when AVX512 is available. (The `_mm256_cvtepi16_epi8` requires AVX512 support)

## Description `sdp_causal` has larger logits error with short prompt, so use a longer prompt in UT