Add a plugin mechanism for operator overrides

[andrej]

Adds a plugin mechanism that allows users to override individual operators. This allows open-source contributors to experiment with dropping in individual custom kernels while harnessing the existing infrastructure.

The plugin mechanism consists of:

• Common infrastructure, public in this repository: flm_plugin.hpp (plugin loading, flm::plugin_context, the FLM_PLUGIN macro), npu_utils/op_override.hpp (op_override, op_result, op_call) and npu_utils/op_registry.hpp (op_registry and its key matching).
• Operator names, per-model and public: each model's header names the operators it dispatches, e.g. gemma4e_ops::op::dequant_qkv and the gemma4e_ops::key() that composes a layer index with one. A plugin includes
  that header and hooks the keys it wants.
• The binding of those names to the engine's dispatch sites, which lives in the per-model DLL/SO (separate PR in internal repo).
• The plugin itself, example below, compiled as a standalone DLL/SO and loaded at runtime via an environment variable.

Declaring operators is opt-in per model: causal_lm::ops() returns nullptr by default, and Gemma4 E2B/E4B is the only engine that declares any so far.

If FLM_PLUGIN is unset, no operator is overridden and behavior is unchanged from "stock".

An override does not have to serve every call it is bound to. Returning flm::op_result::decline() falls back to the default implementation for that call, so a hook can take only the shapes it handles and leave the rest alone.

Example:

#include <memory>

#include "flm_plugin.hpp"                   // op_override, plugin_context, FLM_PLUGIN
#include "models/gemma4e/gemma4e_npu.hpp"   // gemma4e_ops::key, gemma4e_ops::op

class my_dequant : public flm::op_override {
public:
    flm::op_result create_run(const flm::op_call& call) override {
        // ...
    }
};

void register_my_plugin(const flm::plugin_context& ctx) {
    int layer = 0;
    ctx.ops->override_op(gemma4e_ops::key(layer, gemma4e_ops::op::dequant_qkv),
                         std::make_shared<my_dequant>());
}

FLM_PLUGIN(register_my_plugin)

Then start with:

FLM_PLUGIN=/path/to/my_plugin.so flm serve gemma4-it:e2b

This PR also ships an example plugin, src/plugins/iron_gemm, which demonstrates how to override the dequant and matrix multiplication operators of Gemma4 E2B/E4B prefill. src/plugins/README.md documents it.

[andrej]

The CI failure appears to be pre-existing: same error occurs on main.