feat: implement backward computation for more operators

[Ronsor]

This PR will add backward computations for most operators once completed.

• [x] Tanh
• [x] Sigmoid
• [x] GELU + GELU (quick)
• [x] ELU
• [x] clamp
• [x] LeakyReLU
• [x] mean
• [x] concat

Leaving pad, im2col, and norm for a future PR now.

Currently unsure if I should fuse the multiply + gradient computation for gelu_back/gelu_quick_back like with silu_back.