recipe(multilingual-e5-large): add CPU fp32 and fp16 recipes

[codykk]

Summary

Add verified CPU execution provider coverage recipes for intfloat/multilingual-e5-large (feature-extraction, XLM-RoBERTa) with fp32 and fp16 precision. No source changes — current main already builds and runs this model recipe-free; these recipes record exact verified (cpu, cpu, fp32) and (cpu, cpu, fp16) coverage.

Model metadata

• What the model does: Multilingual text embedding model producing dense 1024-dim vectors for semantic similarity, retrieval, and clustering across 100+ languages (confidence: verified — HuggingFace model card + architecture inspection)
• Primary user stories: User supplies multilingual text queries/passages to obtain normalized embeddings for cross-lingual semantic search and sentence similarity
• Supported tasks: feature-extraction (sentence-similarity via cosine; Optimum vendor-registered for xlm-roberta)
• Model architecture:
  XLMRobertaModel (560M params)
  ├── XLMRobertaEmbeddings (word + position + type + LayerNorm)
  └── XLMRobertaEncoder
  └── XLMRobertaLayer ×24 (self-attention + intermediate + output)

Validation and support evidence

Baseline (origin/main 67169d45, winml 0.2.0):

• winml build -m intfloat/multilingual-e5-large → ✅ PASS (82.7s)
• winml perf → ✅ avg 411.65ms, 2.43 samples/s, +1309.1 MB RSS
• winml eval --task feature-extraction → ✅ cosine_spearman 84.0374 (mteb/stsbenchmark-sts, 100 samples)
• Goal floor: L3 (baseline already reaches task metric)

Committed tiers: Effort L0 / Goal L3 / Outcome L0

Per-tuple results:

EP	Device	Precision	L0 (build)	L1 (perf)	L3 (eval)
cpu	cpu	fp32	✅ 94.9s	✅ avg 506.25ms, 2.14 samples/s, +1308.7 MB RSS	✅ cosine_spearman 84.0374 (mteb/stsbenchmark-sts, 100 samples)
cpu	cpu	fp16	✅ 94.3s	✅ avg 445.69ms, 2.01 samples/s, +1308.3 MB RSS	✅ cosine_spearman 84.0374 (mteb/stsbenchmark-sts, 100 samples)

Delta (recipe vs winml config --ep cpu --device cpu):

• fp32: identical to auto-config (filed for verified cpu/cpu coverage)
• fp16: adds quant block (mode: fp16, fp16_keep_io_types: true)

Analyze: CLI-BLOCKED — runtime rule parquet files not found on this host.

Reproduce commands:

# fp32
winml build -m intfloat/multilingual-e5-large \
-c examples/recipes/intfloat_multilingual-e5-large/cpu/cpu/feature-extraction_fp32_config.json \
-o out_fp32 --ep cpu --device cpu --rebuild
winml perf -m out_fp32/model.onnx --ep cpu --device cpu
winml eval -m out_fp32/model.onnx --model-id intfloat/multilingual-e5-large --task feature-extraction --ep cpu --device cpu

# fp16
winml build -m intfloat/multilingual-e5-large \
-c examples/recipes/intfloat_multilingual-e5-large/cpu/cpu/feature-extraction_fp16_config.json \
-o out_fp16 --ep cpu --device cpu --rebuild
winml perf -m out_fp16/model.onnx --ep cpu --device cpu
winml eval -m out_fp16/model.onnx --model-id intfloat/multilingual-e5-large --task feature-extraction --ep cpu --device cpu