ALiBi for the non-flash code path

[Markus28]

Currently, ALiBi is only implemented for the flash code-path. This PR implements ALiBi also for the non-flash code path.

Here is a short script to verify that the behaviour matches the behaviour of the implementation with flash attention. To run this, one needs to merge the changes from #846

import torch
from flash_attn.modules.mha import get_alibi_slopes , FlashSelfAttention, SelfAttention

CAUSAL = False
qkv = torch.randn((16, 55, 3, 4, 32)).to('cuda', torch.float16)

module = SelfAttention(alibi_slopes=torch.tensor(get_alibi_slopes(4)), causal=CAUSAL).to('cuda', torch.float16)
module_flash = FlashSelfAttention(alibi_slopes=torch.tensor(get_alibi_slopes(4)), causal=CAUSAL).to('cuda', torch.float16)
result = module(qkv)
result_flash = module_flash(qkv)
print((result - result_flash) / result_flash * 100)