set tensor names on block params (so weights aren't unnamed leaf_N)

[RapidMark]

GGMLBlock::get_param_tensors already builds each weight's full hierarchical name (e.g. double_blocks.0.img_attn.qkv.weight) but only uses it as the map key — the ggml tensor's own name is never set, so in the compute graph the weight tensors are unnamed and ggml auto-assigns leaf_N.

That makes a couple of things harder than they need to be:

• Graph dumps / profiling show leaf_N instead of the layer name, so you can't tell which weight a node belongs to.
• Tooling that inspects weights by name can't — e.g. importance-matrix / quantization calibration that keys per-weight statistics by the weight name.

The fix is one line, where the full name is already in hand:

for (auto& pair : params) {
    ggml_tensor* param           = pair.second;
    tensors[prefix + pair.first] = pair.second;
    ggml_set_name(param, (prefix + pair.first).c_str());   // added
}

Pure metadata — ggml_set_name only sets the tensor's name field, so there's no change to data, shapes, or the compute graph; generation output is unchanged. Names are bounded by GGML_MAX_NAME as elsewhere. All blocks route their leaf params through this base method, so every weight gets named.

Tested: clean Vulkan build; a Flux (Krea Q4) generation runs and produces a normal image.