context: ignore zero scale LoRAs when checking sameness

[TimNN]

When passed a set of input LoRAs, set_adapters_lora never stores zero scale LoRAs in the loras field, but adapters_lora_are_same expects the number of inputs to be equal to the number of stored LoRAs. So there will always be a mismatch if zero scale adapters are passed repeatedly.

This causes a server started with --lora-init-without-apply to re-reserve the graph for every token when handling requests with "lora": [{"id": 0, "scale": 0.0}].

Tested:

• Started a server with llama-server --model tmp/ggml-org_stories15M_MOE_stories15M_MOE-F16.gguf --lora tmp/moe_shakespeare15M.gguf --lora-init-without-apply
• Sent a chat request with "lora": [{"id": 0, "scale": 0.0}] via the Web UI
• Results without this patch: The server logs print sched_reserve: reserving ... for every token
• Results with this patch: The server prints sched_reserve: reserving ... just once (or not at all)

The server unit tests still pass, but I don't know how to add a test specifically for this bug (unless I want to rely on inference performance, which seems suboptimal).

Alternative fixes considered:

• Filter the list of adapters in set_adapters_lora before calling adapters_lora_are_same. However, that would require allocating even for the not-modified case, which it seemed better to avoid.
• Filter the list of adapters in the server code. However, this functionality is exposed via the public llama_set_adapters_lora API, so it seemed better to fix this for everyone.

AI Disclaimer: Claude Code was used to identify the problem after describing the symptoms. The patch was authored by a human (me).