Fix support for lower VRAM GPUs

[nicolaschan]

I tried running this model on an RTX 4090 with 24GB of memory but the model does not fit in this. I encountered two issues with the current script which this PR resolves:

• There is an attempt to check for lower VRAM GPUs, but this check happens after the .to("cuda") call loads the weights into VRAM, resulting in an OOM exception there before we can get to the low VRAM check.
• Even with 24GB of VRAM, model offload runs out of memory so we should use sequential offload instead. This will slow down inference but brings memory usage down to ~2GB peak.