Add real-model (Qwen3-8B) KV-cache quantization plots

Signed-off-by: Kai Xu <kaix@nvidia.com>

Author: kaix-nv

examples/watersic_kv_cache/README.md

@@ -0,0 +1,37 @@
+# WaterSIC KV-Cache Quantization Example
+
+Synthetic benchmarks comparing KV-cache quantization methods on controlled
+scenarios with known structure. Reproduces the plots from the WaterSIC paper.
+
+## Methods compared
+
+| Method | Format | Hessian-aware | KL-aware |
+|--------|--------|---------------|----------|
+| RTN NVFP-y | Block-scaled fixed-point | No | No |
+| GPTQ NVFP-y | Block-scaled fixed-point | Yes | No |
+| GPTQ-KL NVFP-y | Block-scaled fixed-point | Yes | Yes |
+| WaterSIC-L2 | Entropy-coded | Yes | No |
+| WaterSIC-KL | Entropy-coded | Yes | Yes |
+
+## Usage
+
+```bash
+# Generate all plots
+python kv_cache_watersic_plots.py plot_all
+
+# Individual plots
+python kv_cache_watersic_plots.py plot_temperature
+python kv_cache_watersic_plots.py plot_powerlaw
+python kv_cache_watersic_plots.py plot_retrieval
+python kv_cache_watersic_plots.py plot_sinks
+python kv_cache_watersic_plots.py plot_sinks_cond
+
+# Text-only sweeps (no matplotlib required)
+python kv_cache_watersic_plots.py temperature
+python kv_cache_watersic_plots.py powerlaw
+python kv_cache_watersic_plots.py retrieval
+python kv_cache_watersic_plots.py sinks
+python kv_cache_watersic_plots.py scaling
+```
+
+Plots are saved to `plots/` directory.