Add Lecture 12 code (flash attention)

[ltoniazzi]

Add code for the wonderful lecture by @t-vi about flash attention.

• Reproduced the algo following the lecture exposition so that now it runs on par with scaled_dot_product_attention (on small tensors, on larger I think it needs better allocation of resources)
• Moved cuda code outside the notebook
• Added a bit of profiling to illustrate the register spilling slowdown, with code to illustrate the spilling slowdown in lecture_012/flash_attention_spilling_from_registers.cu and a main.cu to run ncu profiling (discussed in the notebook)