[PWGLF] Add task to produce binned information about V0s/cascades

[romainschotter]

Add task running over strangeness derived data and producing binned information about V0s/cascades

@miranov25 for your information

[github-actions]

O2 linter results: ❌ 62 errors, ⚠️ 0 warnings, 🔕 0 disabled

[alibuild]

Auto-approving on behalf of @romainschotter.

[miranov25]

Thank you for your code, @romainschotter

As discussed on Mattermost and in person, I have some comments on configuration and implementation and would like to share the reasoning behind the strategy I’ve successfully applied for multidimensional group-by statistics extraction since Run 1 (and which contributed to the development of THn in ROOT about 15 years ago).

• I’m aware the memory limit is imposed by ROOT - imposed by buffer size in streames; this is why I previously suggested using THnS. In practice, we will likely use a set of THnS objects, as I do in my performance parameterization code.

Configurable Binning
The variables and binning should remain fully configurable, as the current defaults look too coarse in some cases:

• Invariant mass: The bin width is roughly
  $$\frac{0.547 - 0.447}{25} \approx 0.004$$
  This is too coarse for precise fits. A normalized bin width in terms of $p_T$ resolution would be more appropriate.

• Radius: Binning up to 250 cm should be reduced to match the actual fiducial volume accessible.
  We will prototype improved functional binning with Rahul and Marian in our standalone code for the derived AO2D data
  . For example:

• For φ, we often use coarse global φ bins or local sector φ bins to capture edge-related effectssqrtSaling for pt.
  Once ready we can chack how to post it to the O2 code

---

Why Functional Transformations Matter
Rather than manually specifying non-uniform bins, I recommend using smooth variable transformations.
In my previous multidimensional analyses, the steps were:

1. Apply a functional transformation (e.g., sqrt, asinh) to flatten the distribution or ensure minimal bin width at high $p_T$.
2. Perform group-by statistics (e.g., Gaussian fits) on the transformed grid.

---

Dealing with Sparse Bins
When working in high dimensions, statistics per bin decrease rapidly - 1/nBins. To mitigate this, I used a moving average window across neighboring bins:

• Compute bin centers (COG) dynamically.
• For smoothly varying properties, combine data from ±1 bin per dimension.
  This increases effective statistics by roughly $3^{\text{ndims}}$.

This methodology allowed me to generate performance maps for Run 1, Run 2, and early Run 3. The implementation is currently in AliRoot, but I’m considering porting it to O2, as mixing O2 and AliRoot in the same workflow caused some issues recently - I assume then name clashes for some code - if they used the same variable names in AliRoot and O2.