fix(BA-3308): Support multi-GPU fractional allocation in anti-fragmentation guard

[seedspirit]

resolves #275 (BA-3308)

Checklist: (if applicable)

• Milestone metadata specifying the target backport version
• Mention to the original issue
• Installer updates including:
  • Fixtures for db schema changes
  • New mandatory config options
• Update of end-to-end CLI integration tests in ai.backend.test
• API server-client counterparts (e.g., manager API -> client SDK)
• Test case(s) to:
  • Demonstrate the difference of before/after
  • Demonstrate the flow of abstract/conceptual models with a concrete implementation
• Documentation
  • Contents in the docs directory
  • docstrings in public interfaces and type annotations

[Copilot]

Pull request overview

This PR updates the agent’s fractional allocation anti-fragmentation guard to allow multi-device fractional GPU allocations (BA-3308 / issue #275), and adds unit tests intended to validate the revised behavior.

Changes:

• Revise FractionAllocMap.ensure_slot_not_fragmented() to evaluate multi-device feasibility using per-device “density” and quantum rounding.
• Add an extensive unit-test matrix for density quantization and expected device usage under FILL/EVENLY strategies (plus occupied-device scenarios and edge cases).
• Add a changelog entry describing the fix.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
src/ai/backend/agent/alloc_map.py	Reworks the anti-fragmentation guard to support multi-device fractional allocations.
tests/unit/agent/test_alloc_map.py	Adds new test suites/cases covering defrag density math, strategies, occupied devices, and edge cases.
changes/10477.fix.md	Documents the bugfix in the changelog.

Review notes (blockers):

• The new guard explicitly assumes homogeneous per-device capacity, but the codebase can produce heterogeneous DeviceSlotInfo.amount values (e.g., mock accelerator’s _get_share_raw() varies by device). This can cause false rejections of otherwise feasible allocations.
• The guard’s “remainder/quantum” reasoning is described as matching distribute_evenly, but distribute_evenly operates in self.digits (0.01) while final rounding uses quantum_size (often 0.1 for CUDA shares). This mismatch can allow allocations that later get truncated by round_down(..., quantum_size) such that the returned allocation sum no longer equals the requested amount; the newly added tests currently don’t assert “sum allocated == requested” for the parametrized FILL/EVENLY cases.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[fregataa]

This task does not address the original issue. (Please also update the PR description, it is not related to #275)
It only changes a ensure_slot_not_fragmented() function that checks whether a given agent has enough resources with no fragment.
We have to update the _allocate_by_filling / _allocate_evenly allocator functions.