[kunlunxin] register new ops and update backend configs

[RobertLuobo]

• Register new ops in FULL_CONFIG (atanh, outer, nextafter.out, unbind_copy.int, bitwise_right_shift.Tensor, rnn_relu.input) and add Quantized dispatch key for dequantize.self
• Drop cummin/sort from CUSTOMIZED_UNUSED_OPS
• Add benchmark shapes (grouped_mm, conv_transpose2d, nanmedian, per_token_group_quant_fp8) and baddbmm tune config
• Add gather_block_quantized BLOCK_SIZE_N heuristic (fixes N=0 arange compile error) and tune index_add/index_select heuristics
• black formatting fix in pointwise_dynamic.py

PR Category

Type of Change

Description

Issue

Progress

• Change is properly reviewed (1 reviewer required, 2 recommended).
• Change is responded to an issue.
• Change is fully covered by a UT.

Performance

[Caeruleann]

Registering dequantize.self on the Quantized dispatch key makes per-channel quantized tensors reach this implementation, but it unconditionally calls the per-tensor-only q_scale() and q_zero_point() APIs, and the kernel applies a single scale/zero point to every element. Please support per-channel qschemes using q_per_channel_scales(), q_per_channel_zero_points(), and q_per_channel_axis(), or preserve the native path for unsupported qschemes. Please also add torch.quantize_per_channel parity tests for multiple axes.