Refine XIC v1 to Symbolic Extension Only (No GPU Code)

Removed GPU-related code per specification:
- Deleted xic_extensions/gpu_runtime.py
- Removed GPU logic from op_RUN_PROMPT and op_STREAM
- Removed demo_gpu.gx.json

Kept pure symbolic extension:
- 5 new instructions: STREAM, CHAIN, CALL_GLYPH, SET_CONTEXT, LOG
- Symbolic execution mode via SET_MODE "symbolic"
- run_symbolic_prompt() integration with LAIN cognition layer
- demo_symbolic.gx.json for testing

Implementation now focuses exclusively on:
- Extending instruction set (9 total ops)
- Adding symbolic routing to cognition layer
- Preserving backward compatibility (zero breaking changes)
- No external GPU dependencies

All validation tests pass:
✅ OP_TABLE coverage (9 operations)
✅ XICContext.symbolic_mode field
✅ run_symbolic_prompt() callable
✅ Backward compatibility (demo_chat unchanged)
✅ Symbolic mode execution (LAIN pipeline)
✅ SET_CONTEXT, CHAIN, RUN_PROMPT routing

Constraints met:
✅ No breaking changes
✅ No XIC v2 binary format
✅ No GPU-related code
✅ Strict v1 JSON + .gx architecture

programs/demo_gpu.gx.json

@@ -1,15 +0,0 @@
-{
-  "magic": "GXIC1",
-  "version": 1,
-  "model": "programs/hello_model.gx",
-  "entrypoint": "main",
-  "symbols": {
-    "main": 0
-  },
-  "instructions": [
-    { "op": "LOAD_MODEL", "args": ["programs/hello_model.gx"] },
-    { "op": "SET_PARAM", "args": ["use_gpu", true] },
-    { "op": "LOG", "args": ["Attempting GPU-accelerated execution"] },
-    { "op": "RUN_PROMPT", "args": ["Hello from XIC with GPU acceleration."] }
-  ]
-}