Docs: legacy performance-tuning guide has broken image links and dead anchor; kv_cache_manager.md links removed code

Three docs issues found while reading the current docs tree:

1. docs/source/legacy/performance/performance-tuning-guide/tuning-max-batch-size-and-max-num-tokens.md + deciding-model-sharding-strategy.md: seven image links use ../../media/…, which resolves to docs/source/legacy/media/ — the SVGs actually live in docs/source/media/. All seven are broken on the rendered site. The same tuning guide also links #understanding-the-trt-llm-scheduler, which matches no heading.

2. docs/source/legacy/reference/memory.md:127: links #activation-size, but the only matching heading is ### 2. Activation size → #2-activation-size.

3. docs/source/torch/kv_cache_manager.md:54-55: links tensorrt_llm/_torch/pyexecutor/backend_registries/pytorch_model_registry.py, which no longer exists, and instructs readers to look at create_pytorch_model_based_executor, a function that no longer exists in the code (the KVCacheManager is now instantiated in tensorrt_llm/_torch/pyexecutor/resource_manager.py).

Happy to send PRs for 1-2; for 3 the correct replacement target should be confirmed by a maintainer first.

Start with the two legacy performance-tuning Markdown files, docs/source/legacy/reference/memory.md, and docs/source/torch/kv_cache_manager.md; inspect the referenced headings, image paths, and Python entry points. Verify that all seven images and anchors resolve, then confirm with a maintainer which current resource-manager target should replace the removed registry link and function reference.