Add a Mojo GPU port (CUDA + Metal, Mojo 1.0) to notable forks

[ulmentflam]

This adds one line under the existing Mojo entry in notable forks, following
the pattern of #332.

llm.🔥 is my port of llm.c that
extends @dorjeduck's already-listed Mojo port (Mojo 25.5, CPU) with hand-written
GPU training kernels for Mojo 1.0.0b3: a CUDA/cuBLAS backend, an Apple Metal
backend, and a portable GPU path. What it adds beyond the existing entry: on a
GB10 (DGX Spark) the bf16 path is at parity with llm.c's CUDA trainer (135.97
vs 135.77 ms/step median, B=4, T=1024, GPT-2 124M), and on an M4 Max the Metal
path trains 1.72x faster than PyTorch MPS bf16 (llm.c has no Metal port, so
MPS is the baseline there). Correctness is gated by a ported test_gpt2.mojo:
16 gradient tensors plus a 10-step loss trajectory checked against PyTorch.

Benchmark methodology and reproduction: https://github.com/ulmentflam/llm.mojo#benchmarks

dorjeduck's entry is kept as is; his port is the lineage this builds on.