perf: remove sentence sorting from embedding pipeline

The _order_invariant_text() function previously sorted sentences
alphabetically before embedding to achieve order-invariant embeddings
for the REORDER scenario. This is no longer needed because:

1. Full-document segmented mean pooling (added in v0.2.0) is inherently
   near-order-invariant (0.996 cosine for reordered documents)
2. Sorting broke on code, markdown, and non-English text (split on ". ")
3. Sorting slightly hurt cross-instruction similarity (-0.002)
4. Sorting destroyed instruction-document alignment in long contexts

Benchmark results (SGLang + SemBlend, Qwen2.5-7B-Instruct, A10G):

| Dataset       | WITH sorting | WITHOUT sorting | Delta    |
|---------------|-------------|-----------------|----------|
| TriviaQA      | 3.5%        | 22.6%           | +19.1pp  |
| SCBench       | 4.0%        | 13.6%           | +9.6pp   |
| WikiText103   | 10.0%       | 15.7%           | +5.7pp   |
| LongEval      | 9.7%        | 15.2%           | +5.5pp   |
| NarrativeQA   | 16.7%       | 17.4%           | +0.7pp   |

Hit rates improved on every dataset. Average improvement: +8.1pp.

Signed-off-by: Zach Bennett <zach@worldflowai.com>

Author: zbennett10