refactor: push zero-length normalization into nixlSecDescList (#1752)

## What?
This PR moves the zero-length normalization logic for file-like segments
(`FILE_SEG`, `BLK_SEG`, `OBJ_SEG`) one layer deeper into
`nixlSecDescList`, as suggested in #1551.


## Why?
This is a follow-up refactor to #1551. In that PR the normalization was
applied at the call sites (`addDescList` for writes, `getIndex` for
queries). The
[comments](https://github.com/ai-dynamo/nixl/pull/1551#discussion_r3148738185)
suggested that this invariant should be kept by `nixlSecDescList`
itself, and it would be fail-safe to not assume that the input is
already given normalized.


## How?
- **Write path**: `addDesc()` copies the input, normalizes it if the
list type is unbounded, then inserts. `addDescs()` applies
`normalizeBatch()` on the whole vector before sorting and merging.
- **Query path**: `getIndex()` still normalizes the query via
`normalizeQuery()`, because externally-provided queries cannot be
assumed to be pre-normalized.



<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Refactor**
* Centralized section descriptor normalization inside the
section-descriptor list to ensure consistent handling when adding and
querying descriptors.
* **Bug Fixes**
* Improved matching behavior for lookups by aligning how queries are
normalized with how stored descriptors are ordered.
* Adjusted behavior so only designated segment types treat `len=0` as
unbounded (normalized to max length), while others preserve `len=0`.
* **Tests**
* Added unit tests covering zero-length descriptor handling across
segment types and add/move/query scenarios.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Co-authored-by: Ilia Yastrebov <iyastrebov@nvidia.com>