[single] Refactor setting tensors info in ml_single_open_custom

[songgot]

Improve ml_single_open_custom's framework-specific in/out tensors info settings handling to a table-based approach.
Improve code readability.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR replaces the if (nnfw == TF || SNAP || PYTORCH || TRIX || NCNN) ... else if (nnfw == ARMNN) block in ml_single_open_custom with a static const table (requires_in_info / requires_out_info) plus a lookup loop. I compared every path against upstream/main (whose copy of this block is identical to the PR base; git merge-tree is clean): required-group frameworks still fail with ML_ERROR_INVALID_PARAMETER before any property is set when either info is missing, in/out are still set in the same order via the same helper, ARMNN still treats both as optional, and the resolved local nnfw is used as before. Behavior is equivalent; only log text changes. Non-blocking comments below.

1. [Low] gst-indent violations — c/src/ml-api-inference-single.c:1154, 1157-1158, 1164, 1166-1168, 1173: over-long status = ml_single_set_inout_tensors_info (...) lines (original wrapped after status =), continuation indented 14 instead of 12 spaces, _ml_error_report_continue misaligned at 1166, _ml_logi("...") missing the GNU space before ( and >100 columns, and two consecutive blank lines at 1043-1044. Note indent.sh in CI only emits warnings, so this must be fixed manually with the repo's indent flags.
2. [Low] New _ml_logi fires on the most common path — :1172-1174: every framework not in the table (tflite, nnfw, snpe, onnx, tvm, tensorrt, qnn, llama.cpp, custom, ...) now logs "not found in tensors info table, using default behavior" on every open, which reads as if something is missing. Drop the else branch (matches the original) or downgrade to _ml_logd with wording like "no tensors-info requirement for '%s'".
3. [Low] ARMNN diagnostics lost specificity — :1156-1158, 1166-1168: the old ARMNN branch mentioned that infos are optional and included fw_name; both branches now share a generic message. Return codes are unchanged. Suggest including fw_name in the shared message.
4. [Low] Naming / Doxygen — :1018-1042: typedef nnfw_tensors_info lacks the repo's _s suffix, and the struct doc uses @param for members (Doxygen will warn); document members inline with /**< ... */.
5. [Low] Readability gain is small — the table encodes only two cases (both required / both optional) with no asymmetric entries, yet adds a struct, 6 rows, three locals, a loop, and a found_in_table flag (+21 lines). A small helper returning an enum would achieve the same with less machinery; fine if asymmetric requirements are anticipated.
6. [Low] Overlap with #651 — your open #651 extracts this exact region into helpers; whichever merges second will conflict. Consider folding this change into #651 or rebasing one on the other explicitly.
7. [Info] Test coverage gap (pre-existing) — no unit test reaches the "required info missing -> ML_ERROR_INVALID_PARAMETER" path with a valid model, so a slip in this block would pass CI. Optional follow-up: a test opening a valid model with a required-info nnfw and NULL infos, skipped when the framework is unavailable.

No back-door or suspicious behavior found.